Academia.eduAcademia.edu

Figure 145 – uploaded by Carlos B Westphall

See full PDFdownloadDownload figure
Figure 4. Overall error rates for the Skeys-MAP, MsWord, and iSpell algorithms, when considering all error types and only substitution errors, respectively In Table II, we show the initial error distribution and how many errors of each type were not corrected by the algorithm. It can be noticed that the correction method is less efficient for deletions, when a word has been split, and for combined error types. In the case of deletions, the typed

Figure 4 Overall error rates for the Skeys-MAP, MsWord, and iSpell algorithms, when considering all error types and only substitution errors, respectively In Table II, we show the initial error distribution and how many errors of each type were not corrected by the algorithm. It can be noticed that the correction method is less efficient for deletions, when a word has been split, and for combined error types. In the case of deletions, the typed

Related Figures (171)

The physical representation of our model is illustrated in Figure 1. The convection of air from air supply to the air return vent near the ceiling is represented by the PDE part. The diffusive term is intentionally omitted since it plays a relatively minor role in dispersing indoor pollutants as suggested in [10]. Another design consideration involved is the modeling of the COz concentration near the ceiling since this is where we see most effect from human-generated CO. One explanation is that the warm breath from a human occupant acts as a “bubble” of gas that rises to the ceiling, since it is more buoyant than the ambient, cooler air. Thus, the air coming from lower in the room is modeled as a source term on the PDE across its entire path. The fact that this bubble of air does not immediately rise to the ceiling but only gradually (as observed in the response of the CO, concentration in the room due to changes of the human’s COz2 input shown in Figures 5 and 8, during the occupant experiments) is captured by the ODE part of the model, which behaves as a filter between the unknown CO, emission rate of humans and the CO. concentration in the room. Figure 1. The physical representation of the model. Fresh air with CO2 concentration U(t) enters the room from the supply vent, and exits the room after convection and mixing with human breath, V(t), which rises to the ceilings, and the measured COz2 concentration at the return vent is u(1, t).
The physical representation of our model is illustrated in Figure 1. The convection of air from air supply to the air return vent near the ceiling is represented by the PDE part. The diffusive term is intentionally omitted since it plays a relatively minor role in dispersing indoor pollutants as suggested in [10]. Another design consideration involved is the modeling of the COz concentration near the ceiling since this is where we see most effect from human-generated CO. One explanation is that the warm breath from a human occupant acts as a “bubble” of gas that rises to the ceiling, since it is more buoyant than the ambient, cooler air. Thus, the air coming from lower in the room is modeled as a source term on the PDE across its entire path. The fact that this bubble of air does not immediately rise to the ceiling but only gradually (as observed in the response of the CO, concentration in the room due to changes of the human’s COz2 input shown in Figures 5 and 8, during the occupant experiments) is captured by the ODE part of the model, which behaves as a filter between the unknown CO, emission rate of humans and the CO. concentration in the room. Figure 1. The physical representation of the model. Fresh air with CO2 concentration U(t) enters the room from the supply vent, and exits the room after convection and mixing with human breath, V(t), which rises to the ceilings, and the measured COz2 concentration at the return vent is u(1, t).
Figure 2. (a) CO sensor up close. The platform is integrated with the main module to sense COz concentration, and a local data storage solution with SD card. (b) The testbed is a conference room located at the Center for Research in Energy Systems Trasnformation (CREST) in Cory Hall on the UC Berkeley campus. The space is of size 14 x 10 x 9 ft?, equipped with a full ventilation system including an air return vent and air supply vent, as illustrated in Figure 1.
Figure 2. (a) CO sensor up close. The platform is integrated with the main module to sense COz concentration, and a local data storage solution with SD card. (b) The testbed is a conference room located at the Center for Research in Energy Systems Trasnformation (CREST) in Cory Hall on the UC Berkeley campus. The space is of size 14 x 10 x 9 ft?, equipped with a full ventilation system including an air return vent and air supply vent, as illustrated in Figure 1.
Figure 5. Occupants experiment F. Field measurements during project discussion. Top: Proxy measurements. Bottom: Corresponding occupancy. The following shows results from several such experiments, which substantially cover the usage of the conference room, and can be easily extended to other areas in the building. The field measurements are shown in Figure 5 and 8 (right), and the strictly controlled experiment is illustrated in Figure 8 (left). Note that to avoid significant overlap between graphs of this section and those of the simulation section, we arbitrarily decide which graphs show the blackboard measurement and the others show the simulated return, as long as the evidence is sufficient to for the argument.
Figure 5. Occupants experiment F. Field measurements during project discussion. Top: Proxy measurements. Bottom: Corresponding occupancy. The following shows results from several such experiments, which substantially cover the usage of the conference room, and can be easily extended to other areas in the building. The field measurements are shown in Figure 5 and 8 (right), and the strictly controlled experiment is illustrated in Figure 8 (left). Note that to avoid significant overlap between graphs of this section and those of the simulation section, we arbitrarily decide which graphs show the blackboard measurement and the others show the simulated return, as long as the evidence is sufficient to for the argument.
Figure 4. CO2 pump experiment B. Short term excitation with period of 1 hour. Measurements at return (red), supply (green) vents, and blackboard (blue) are shown. Figure 3. CO2 pump experiment D. The measured CO2 concentration from different locations for a 5-hour CO2 release are shown.
Figure 4. CO2 pump experiment B. Short term excitation with period of 1 hour. Measurements at return (red), supply (green) vents, and blackboard (blue) are shown. Figure 3. CO2 pump experiment D. The measured CO2 concentration from different locations for a 5-hour CO2 release are shown.
Though the system is responsive to the change of occu- pancy, the time it takes to accumulate or deplete CO, to the corresponding stationary value is fairly long. From vacancy to a high level occupancy, the measurement slowly sweeps across several intermediate levels. The difficulty of most distribution- based classification methods is illustrated in Figure 6, where the significant overlapping of regions and misplacement of modes corresponding to different levels of occupancy will lead to confusion for standard machine learning algorithms. By modeling the temporal and spatial dynamics of the system, as we demonstrate in the subsequent sections, we can develop an inference method that is both robust to noise and responsive to change of occupancy. Figure 6. Empirical distribution of CO2 concentration for all occupants experiments corresponding to different occupancy (color coded). B. Simulation with Proxy Link Model
Though the system is responsive to the change of occu- pancy, the time it takes to accumulate or deplete CO, to the corresponding stationary value is fairly long. From vacancy to a high level occupancy, the measurement slowly sweeps across several intermediate levels. The difficulty of most distribution- based classification methods is illustrated in Figure 6, where the significant overlapping of regions and misplacement of modes corresponding to different levels of occupancy will lead to confusion for standard machine learning algorithms. By modeling the temporal and spatial dynamics of the system, as we demonstrate in the subsequent sections, we can develop an inference method that is both robust to noise and responsive to change of occupancy. Figure 6. Empirical distribution of CO2 concentration for all occupants experiments corresponding to different occupancy (color coded). B. Simulation with Proxy Link Model
Figure 7. Proxy model simulation with COz pump experiment A with 30 minutes (left) and experiment B with | hour (right) periodic excitation. Measurements at supply (green), return (blue), and simulated return (red) vents are shown.
Figure 7. Proxy model simulation with COz pump experiment A with 30 minutes (left) and experiment B with | hour (right) periodic excitation. Measurements at supply (green), return (blue), and simulated return (red) vents are shown.
Figure 8. Proxy link model simulation with occupants experiment E (left) and G (right). The proxy measurements at return (blue), supply (green), and simulated return (red) vents are demonstrated.
Figure 8. Proxy link model simulation with occupants experiment E (left) and G (right). The proxy measurements at return (blue), supply (green), and simulated return (red) vents are demonstrated.
Figure 10. Sensing the latent CO2 emission rate by proxy for CO2 pump experiment A (left) and experiment C (right). The estimated emission rate (blue), median filtered rate (green), and ground truth (red) are given.
Figure 10. Sensing the latent CO2 emission rate by proxy for CO2 pump experiment A (left) and experiment C (right). The estimated emission rate (blue), median filtered rate (green), and ground truth (red) are given.
Figure 9. Spatial and temporal dynamics of CO2 concentration as represented as the states in the proxy link model.
Figure 9. Spatial and temporal dynamics of CO2 concentration as represented as the states in the proxy link model.
TABLE IV. COMPARISON OF ROOT MEAN-SQUARED ERROR OF ESTIMATION IN OCCUPANTS EXPERIMENTS to yield different outputs due to different training, sensing by proxy will output the same value given the chosen parameters, which is desirable since it is less susceptible to training noise.
TABLE IV. COMPARISON OF ROOT MEAN-SQUARED ERROR OF ESTIMATION IN OCCUPANTS EXPERIMENTS to yield different outputs due to different training, sensing by proxy will output the same value given the chosen parameters, which is desirable since it is less susceptible to training noise.
Figure 11. Occupancy detection by “sensing by proxy” (Algorithm 1) for experiment E (left) and G (right). The response times from vacancy to occupancy and vice versa are about 10 seconds and 5 to 10 minutes respectively, since the method detects the dynamics of the system rather than static concentration. The estimations (green) are within 1 occupant of the ground truth (red).
Figure 11. Occupancy detection by “sensing by proxy” (Algorithm 1) for experiment E (left) and G (right). The response times from vacancy to occupancy and vice versa are about 10 seconds and 5 to 10 minutes respectively, since the method detects the dynamics of the system rather than static concentration. The estimations (green) are within 1 occupant of the ground truth (red).
Figure 12. Visualization of confusion matrix for Bayes Net (left) and sensing by proxy (right), where the position of circles represents the true number ¢ occupants (x-axis) and estimated number of occupants (y-axis), and the size indicates the percentage normalized for each column.
Figure 12. Visualization of confusion matrix for Bayes Net (left) and sensing by proxy (right), where the position of circles represents the true number ¢ occupants (x-axis) and estimated number of occupants (y-axis), and the size indicates the percentage normalized for each column.
Figure 1. Analysis of papers by publication year. In this study, we identified usability evaluation approaches for mobile apps published from 2004 to 2014 (when the study was run).
Figure 1. Analysis of papers by publication year. In this study, we identified usability evaluation approaches for mobile apps published from 2004 to 2014 (when the study was run).
Figure 3. Analysis of mobile application category Finally, evaluations in hybrid apps are starting to emerge, because it represents a new trend of development of mobile apps, justifying the small number of research in the area. All authors who proposed usability evaluation approaches for hybrid apps highlighted that this category is emerging and needs more research, not only for evaluation of usability, but also for application development.
Figure 3. Analysis of mobile application category Finally, evaluations in hybrid apps are starting to emerge, because it represents a new trend of development of mobile apps, justifying the small number of research in the area. All authors who proposed usability evaluation approaches for hybrid apps highlighted that this category is emerging and needs more research, not only for evaluation of usability, but also for application development.
Figure 2. Analysis by evaluation technique category We observed the category of usability evaluation techniques most frequent are dynamic approaches (77/101 for mobile apps; 25/28 for ubiquitous mobile apps). This result is justified due to the need of assessing the app on a scenario closer to reality, possibly by using dynamic approaches, making the evaluation more efficient.
Figure 2. Analysis by evaluation technique category We observed the category of usability evaluation techniques most frequent are dynamic approaches (77/101 for mobile apps; 25/28 for ubiquitous mobile apps). This result is justified due to the need of assessing the app on a scenario closer to reality, possibly by using dynamic approaches, making the evaluation more efficient.
Figure 4. Analysis by type of experiment.
Figure 4. Analysis by type of experiment.
Figure 6. Analysis of Usability Factors Yet to answer the question Q3 addressed in section III.A, an analysis by usability factors was done, as shown in Figure 6. Analyzing the results, we observed that the factors uset (60/108) and tasks (66/108) are more frequent in the selected papers. In general, most of the papers when dealing with one of these factors also deal with the second one. The evaluation of the factor context of use was observed only in approaches that deal with context awareness requirements. In this analysis, a paper could address one or more factors, justifying that the sum of the numbers distributed among the factors is greater than the number of identified papers. Furthermore, this also shows us that ubiquitous mobile applications can be evaluated recitals three factors as context of use with 21 identified papers, the user featuring 14 papers and 17 papers task with a total of 28.
Figure 6. Analysis of Usability Factors Yet to answer the question Q3 addressed in section III.A, an analysis by usability factors was done, as shown in Figure 6. Analyzing the results, we observed that the factors uset (60/108) and tasks (66/108) are more frequent in the selected papers. In general, most of the papers when dealing with one of these factors also deal with the second one. The evaluation of the factor context of use was observed only in approaches that deal with context awareness requirements. In this analysis, a paper could address one or more factors, justifying that the sum of the numbers distributed among the factors is greater than the number of identified papers. Furthermore, this also shows us that ubiquitous mobile applications can be evaluated recitals three factors as context of use with 21 identified papers, the user featuring 14 papers and 17 papers task with a total of 28.
Figure 5. Usability Attributes Analysis In order to answer the question Q3 addressed in Section IlI.A, an analysis by usability attributes was performed, as shown in Figure 5. We observed that user satisfaction is the most investigated attribute in the identified approaches (67/101 papers) on mobile apps, and it is the second more investigated in ubiquitous mobile apps (15/28 papers). In this analysis, a paper could address one or more attributes, justifying that the sum of the numbers distributed among the attributes is greater than the number of identified papers.
Figure 5. Usability Attributes Analysis In order to answer the question Q3 addressed in Section IlI.A, an analysis by usability attributes was performed, as shown in Figure 5. We observed that user satisfaction is the most investigated attribute in the identified approaches (67/101 papers) on mobile apps, and it is the second more investigated in ubiquitous mobile apps (15/28 papers). In this analysis, a paper could address one or more attributes, justifying that the sum of the numbers distributed among the attributes is greater than the number of identified papers.
Figure 7. Feature of ubiquity analysis
Figure 7. Feature of ubiquity analysis
After that, the execution of scripts was implemented using the functions created, as described in Section III. An example of automation can be seen in Figure 2 that validates the behavior of Waze app. In short, the script defines the expected behavior through image (line 2 and 3) for both location ON and OFF, modifies the status of the location property of the phone (to ON), opens the app and verifies whether Waze has adapted to the new context or not. The final steps were to run the scripts for the remainder of the apps and compare the results after execution.
After that, the execution of scripts was implemented using the functions created, as described in Section III. An example of automation can be seen in Figure 2 that validates the behavior of Waze app. In short, the script defines the expected behavior through image (line 2 and 3) for both location ON and OFF, modifies the status of the location property of the phone (to ON), opens the app and verifies whether Waze has adapted to the new context or not. The final steps were to run the scripts for the remainder of the apps and compare the results after execution.
It used the Windows Phone emulator called Project My Screen App [8] to project the application screen to be possible run the scripts. This is the last step.
It used the Windows Phone emulator called Project My Screen App [8] to project the application screen to be possible run the scripts. This is the last step.
The high-level architecture of CardiaCare service is pre- sented in Figure 1.
The high-level architecture of CardiaCare service is pre- sented in Figure 1.
Figure 2. Using attribute dictionaries
Figure 2. Using attribute dictionaries
Commonly used optimizations can be described with the following procedure illustrated in (Figure 3).
Commonly used optimizations can be described with the following procedure illustrated in (Figure 3).
Figure 5. Separate tables for attributes of different types Figure 4. One table for all attribute values
Figure 5. Separate tables for attributes of different types Figure 4. One table for all attribute values
Figure 6. Hirarchical dictionaries in CardiaCare data model
Figure 6. Hirarchical dictionaries in CardiaCare data model
Figure 7. Attribute representation in CardiaCare data model
Figure 7. Attribute representation in CardiaCare data model
Figure 1. Previous Projects Developed in Urban Computing. UrbanContext was designed from the result of previous projects and experiments developed in controlled and open environments of urban interaction. Our previous experiences have been performed gradually and were restricted to three major projects [14][15][16] (See Figure 1).
Figure 1. Previous Projects Developed in Urban Computing. UrbanContext was designed from the result of previous projects and experiments developed in controlled and open environments of urban interaction. Our previous experiences have been performed gradually and were restricted to three major projects [14][15][16] (See Figure 1).
Figure 2. UrbanContext Model.
Figure 2. UrbanContext Model.
Figure 4. UrbanContext Android Application.
Figure 4. UrbanContext Android Application.
Figure 1 illustrates the architecture of adaptive streaming for a virtualization service. In this architecture, HTTP 2.0 is used for server push. The proposed system performs segment quality control and HTTP request control algorithm to The rest of this paper is organized as follows: in Section II, we describe the concepts and algorithms introduced in the proposed scheme. In Section III, we show the simulation results. Conclusions and future works are presented in Section IV. To reduce the latency of DASH, the server push based streaming scheme has been proposed. The server push mechanism pushes a resource directly to the client without the client request. However, server push is not suitable for DASH as the DASH adjusts the bitrate of contents based on the client request. In this paper, we propose a DASH based adaptive streaming scheme to improve the QoS of virtualization service. The proposed scheme provides seamless display by adjusting the quality of the segment based on the segment throughput and the buffer status. It also reduces latency by using the server push mechanism.
Figure 1 illustrates the architecture of adaptive streaming for a virtualization service. In this architecture, HTTP 2.0 is used for server push. The proposed system performs segment quality control and HTTP request control algorithm to The rest of this paper is organized as follows: in Section II, we describe the concepts and algorithms introduced in the proposed scheme. In Section III, we show the simulation results. Conclusions and future works are presented in Section IV. To reduce the latency of DASH, the server push based streaming scheme has been proposed. The server push mechanism pushes a resource directly to the client without the client request. However, server push is not suitable for DASH as the DASH adjusts the bitrate of contents based on the client request. In this paper, we propose a DASH based adaptive streaming scheme to improve the QoS of virtualization service. The proposed scheme provides seamless display by adjusting the quality of the segment based on the segment throughput and the buffer status. It also reduces latency by using the server push mechanism.
To evaluate the performance, we have implemented the proposed scheme in a DASH reference player developed by DASH Industry Forum (IF). We compared the amount of playback buffer among the proposed scheme, RAHS, ASAC, and default adaptation algorithm of DASH IF player. Figure 3 shows that the proposed scheme stably maintains the buffer level without buffer underflow because the quality is adjusted on the basis of the segment throughput and buffer status. Figure 3. Comparison of the amount of playback buffer
To evaluate the performance, we have implemented the proposed scheme in a DASH reference player developed by DASH Industry Forum (IF). We compared the amount of playback buffer among the proposed scheme, RAHS, ASAC, and default adaptation algorithm of DASH IF player. Figure 3 shows that the proposed scheme stably maintains the buffer level without buffer underflow because the quality is adjusted on the basis of the segment throughput and buffer status. Figure 3. Comparison of the amount of playback buffer
Figure 2. Behavior of the proposed adaptive streaming sceme Figure 2 depicts the behavior of the proposed adaptive streaming scheme. At the beginning of the service, the client requests virtualized content which has the minimum bitrate because there is no information about the network status and the playback buffer is empty. Then, the server pushes the segments until the client requests another bitrate of virtualized content. Based on this approach, the proposed scheme is better able to reduce latency and prevent playback discontinuity.
Figure 2. Behavior of the proposed adaptive streaming sceme Figure 2 depicts the behavior of the proposed adaptive streaming scheme. At the beginning of the service, the client requests virtualized content which has the minimum bitrate because there is no information about the network status and the playback buffer is empty. Then, the server pushes the segments until the client requests another bitrate of virtualized content. Based on this approach, the proposed scheme is better able to reduce latency and prevent playback discontinuity.
We compared the latency between the proposed scheme with server push and one without server push mechanism. Figure 4 shows that the latency is evidently decreased when the server push mechanism is deployed. DASH can improve the QoS of streaming-based virtualization service by adjusting the quality of content according to the network condition. However, HTTP request/response procedure of DASH introduces additional latency which degrades the quality of virtualization service. In this paper, we propose an adaptive streaming scheme for DASH to improve the QoS of streaming-based virtualization service. The simulation results show that the proposed scheme provides seamless playback and low latency by adjusting the quality of content based on the server push mechanism of HTTP 2.0. In the future work, we will enhance the proposed adaptive streaming scheme to reduce the frequent quality change due to the varying network condition.
We compared the latency between the proposed scheme with server push and one without server push mechanism. Figure 4 shows that the latency is evidently decreased when the server push mechanism is deployed. DASH can improve the QoS of streaming-based virtualization service by adjusting the quality of content according to the network condition. However, HTTP request/response procedure of DASH introduces additional latency which degrades the quality of virtualization service. In this paper, we propose an adaptive streaming scheme for DASH to improve the QoS of streaming-based virtualization service. The simulation results show that the proposed scheme provides seamless playback and low latency by adjusting the quality of content based on the server push mechanism of HTTP 2.0. In the future work, we will enhance the proposed adaptive streaming scheme to reduce the frequent quality change due to the varying network condition.
Figure 1. High-Level Block Diagram of the system parts. The first part is a traditional online social network that enables users to communicate through the use of comments. Furthermore, this social network enables users to create con- nections with other individuals, most importantly, those users that are within their vicinity. The second part of this social network is another website, which we refer to as the Interactive Wall. This website is responsible for extracting a stream of comments posted by the users of our system. This stream of comments is then displayed onto a big screen which is placed in a public space. Apart from displaying the stream of comments, the Interactive Wall will also try to play an active role in the discussions by displaying articles relevant to the current discussion on the wall (which are selected from news sources using artificial intelligence techniques).
Figure 1. High-Level Block Diagram of the system parts. The first part is a traditional online social network that enables users to communicate through the use of comments. Furthermore, this social network enables users to create con- nections with other individuals, most importantly, those users that are within their vicinity. The second part of this social network is another website, which we refer to as the Interactive Wall. This website is responsible for extracting a stream of comments posted by the users of our system. This stream of comments is then displayed onto a big screen which is placed in a public space. Apart from displaying the stream of comments, the Interactive Wall will also try to play an active role in the discussions by displaying articles relevant to the current discussion on the wall (which are selected from news sources using artificial intelligence techniques).
Figure 3. Purposes of using Occupy
Figure 3. Purposes of using Occupy
of previous discussions. The vast majority of the comments posted (94%) were text-only comments, while the remaining 6% included images. Figure 2 shows the usage of our system over a period of time. This includes a day before the actual start of the experiment, and a period of three days after it had ended. Figure 2. Usage of the system over time
of previous discussions. The vast majority of the comments posted (94%) were text-only comments, while the remaining 6% included images. Figure 2 shows the usage of our system over a period of time. This includes a day before the actual start of the experiment, and a period of three days after it had ended. Figure 2. Usage of the system over time
Figure 2. Multimodal localization system fusing WiFi signals and images. Positioning range is easily broaden by simply adding more APs (Access Point) in the environment. (3) Computational expense is much lower for WiFi-based than vision-based positioning system. (4) The positioning accuracy is provided in accordance with the application-specific demands. Although the camera- based system brings more impressed positioning results, but not every where in building needs high localizing accuracy. (5) Sampling frequency is improved for the WiFi-based system, because it has lower sampling rate (about one signal measure per second) than vision-based system (approximately 15 fps). (6) The information for person Re-ID becomes richer. One object can be identified simultaneously by both WiFi and camera systems. These ID cues can be used in the way of supporting one another in multi-model object localization system. For example, at a certain time, one object is localized and identified by WiFi system, with the position of Pyy;p; and the identity of /Dw:r; (the MAC address of mobile device) respectively. At the same time, this object is also determined by Peam and ID-am from camera system. However, Py ir; is not as accurate as P.gm, whilst [Dw ;pr; is clearer than [Deam. Therefore, by using both of these systems, the object can be localized by P.a, and identified by [Dy ;7;.
Figure 2. Multimodal localization system fusing WiFi signals and images. Positioning range is easily broaden by simply adding more APs (Access Point) in the environment. (3) Computational expense is much lower for WiFi-based than vision-based positioning system. (4) The positioning accuracy is provided in accordance with the application-specific demands. Although the camera- based system brings more impressed positioning results, but not every where in building needs high localizing accuracy. (5) Sampling frequency is improved for the WiFi-based system, because it has lower sampling rate (about one signal measure per second) than vision-based system (approximately 15 fps). (6) The information for person Re-ID becomes richer. One object can be identified simultaneously by both WiFi and camera systems. These ID cues can be used in the way of supporting one another in multi-model object localization system. For example, at a certain time, one object is localized and identified by WiFi system, with the position of Pyy;p; and the identity of /Dw:r; (the MAC address of mobile device) respectively. At the same time, this object is also determined by Peam and ID-am from camera system. However, Py ir; is not as accurate as P.gm, whilst [Dw ;pr; is clearer than [Deam. Therefore, by using both of these systems, the object can be localized by P.a, and identified by [Dy ;7;.
Figure 1. Flowchart of object localization system. Object localization is known as a problem of determining the object position in the environment. For each user in multi- user localization system, two problems of positioning (where the user is) and identifying (who the user is) must be solved simultaneously. A general diagram for object localization is illustrated in Figure 1. In this figure, the input cues can come from different sensors, such as optical, radio frequency, ultra sound, inertia, DC Electromagnetic sensors, etc. From the input cues, localization and Re-ID are executed simultaneously to give the output for object position and ID. Multimodal object localization is defined as a problem of multi-cue combination for input or fusion of different positioning methods. As proven by Vinyals et al. [17] and Dao et al. [18], compounding of different models gives better positioning results than applying a single model. Teixeira et al. [19] proposed to use the motion signature taken from wearable accelerometer for identifying people in camera network.
Figure 1. Flowchart of object localization system. Object localization is known as a problem of determining the object position in the environment. For each user in multi- user localization system, two problems of positioning (where the user is) and identifying (who the user is) must be solved simultaneously. A general diagram for object localization is illustrated in Figure 1. In this figure, the input cues can come from different sensors, such as optical, radio frequency, ultra sound, inertia, DC Electromagnetic sensors, etc. From the input cues, localization and Re-ID are executed simultaneously to give the output for object position and ID. Multimodal object localization is defined as a problem of multi-cue combination for input or fusion of different positioning methods. As proven by Vinyals et al. [17] and Dao et al. [18], compounding of different models gives better positioning results than applying a single model. Teixeira et al. [19] proposed to use the motion signature taken from wearable accelerometer for identifying people in camera network.
Figure 3. Surveillance region with WiFi range and disjoint cameras’ FOVs. In the surveillance region, two problems of person local- ization and Re-ID will be solved concurrently by combining visual and WiFi information. Figure 3 demonstrates a surveil- lance region which contains WiFi range and camera FOVs. In this region, the WiFi range covers some visual ranges (the camera FOVs: FOV of C;, FOV of Co,.., FOV of C,,). This means the user always move within WiFi range but switch from one camera FOV to others.
Figure 3. Surveillance region with WiFi range and disjoint cameras’ FOVs. In the surveillance region, two problems of person local- ization and Re-ID will be solved concurrently by combining visual and WiFi information. Figure 3 demonstrates a surveil- lance region which contains WiFi range and camera FOVs. In this region, the WiFi range covers some visual ranges (the camera FOVs: FOV of C;, FOV of Co,.., FOV of C,,). This means the user always move within WiFi range but switch from one camera FOV to others.
Figure 5. The basic idea of representation based on kernel methods. where G is defined by: G'G= Kap and Kpp is D x D matrix with {Kee}i; = k(vi,v;). kp is a D x 1 vector with {ke}; = k(x, uj).
Figure 5. The basic idea of representation based on kernel methods. where G is defined by: G'G= Kap and Kpp is D x D matrix with {Kee}i; = k(vi,v;). kp is a D x 1 vector with {ke}; = k(x, uj).
Figure 7 shows the 8th floor plan of our office building. It is set as the testing environment for our combined person localization system. At the entrance, people hold smart phones or tablets go one by one through the check-in gate, then move inside the surveillance area, and finish their routes by going out check-out gate.
Figure 7 shows the 8th floor plan of our office building. It is set as the testing environment for our combined person localization system. At the entrance, people hold smart phones or tablets go one by one through the check-in gate, then move inside the surveillance area, and finish their routes by going out check-out gate.
Figure 6. Illustration of size-adaptive patches (a, c) and size-fixed patches (a, b) which is mentioned in [1].
Figure 6. Illustration of size-adaptive patches (a, c) and size-fixed patches (a, b) which is mentioned in [1].
ihe dataset for Numan WKe-1V includes 2) people with different routes in the testing environment. Each person spends from 3 to 5 minutes for his route. An approximation of 800 values of RSSIs are scanned, about 2000 frames are captured for each camera in the surveillance area. All cap- tured frames are processed as real Re-ID scenario of mul- timodal pedestrian localization system. Firstly, the images of person body at different poses are extracted from video sequence of the entrance camera. They are later used for training phase of appearance-based person identification. In the surveillance area, each frame from the sequences of four cameras is processed for human detection, so each of 25 pedestrians will have the appearance images (the bounding boxes of each person) at different views. The image file- name identifies the video sequence to which it belongs, the camera ID, frame number, time (hour, minute, second), and the person ID: VVV_WW_XXXXX_YYYYYY_Z22Z.jpg (E.g., 025_01_01260_153702_012.4pg means the appearance image belongs to video sequence 25; it is captured from the camera 01; frame number 1260; the time is 15h:37m:02s; the person ID is 12). These images are utilized for testing phase of person Re-ID system. An example in the dataset for person Re-ID in camera network is shown in Figure 8. The images on the top are used for training phase of appearance-based person identification. They are captured by a camera at the check-in/check-out region. The images for testing phase are shown at the bottom. They come from four different cameras in the surveillance region. Figure 8. The instances in dataset for vision-based person Re-ID.
ihe dataset for Numan WKe-1V includes 2) people with different routes in the testing environment. Each person spends from 3 to 5 minutes for his route. An approximation of 800 values of RSSIs are scanned, about 2000 frames are captured for each camera in the surveillance area. All cap- tured frames are processed as real Re-ID scenario of mul- timodal pedestrian localization system. Firstly, the images of person body at different poses are extracted from video sequence of the entrance camera. They are later used for training phase of appearance-based person identification. In the surveillance area, each frame from the sequences of four cameras is processed for human detection, so each of 25 pedestrians will have the appearance images (the bounding boxes of each person) at different views. The image file- name identifies the video sequence to which it belongs, the camera ID, frame number, time (hour, minute, second), and the person ID: VVV_WW_XXXXX_YYYYYY_Z22Z.jpg (E.g., 025_01_01260_153702_012.4pg means the appearance image belongs to video sequence 25; it is captured from the camera 01; frame number 1260; the time is 15h:37m:02s; the person ID is 12). These images are utilized for testing phase of person Re-ID system. An example in the dataset for person Re-ID in camera network is shown in Figure 8. The images on the top are used for training phase of appearance-based person identification. They are captured by a camera at the check-in/check-out region. The images for testing phase are shown at the bottom. They come from four different cameras in the surveillance region. Figure 8. The instances in dataset for vision-based person Re-ID.
Figure 9. The results of proposed method against AHPE [23], SDALF [24] and KDES [1] on (a) CAVIAR4REID dataset and (b) iLIDS dataset. We compare the results of our proposed method with orig- inal KDES [1] and other state-of-the-art approaches on multi- shot datasets of CAVIAR4REID and iLIDs. The experimental settings are kept the same as in [23], with modified version of iLIDS dataset is selected (including only 69 individuals, with at least 4 images for each) and 72 pedestrians for CAVIAR4REID dataset, in which 50 different individuals are captured under both views. Each view has 10 images for each pedestrian. The outperforming results of the proposed method are shown in Figure 9. For CAVIAR4REID dataset, rank-1 recognition rate of AHPE (Asymmetry-based Histogram Plus Epitome) [23] is much lower than our method. It is only about 8 %, compared with 67.76 % of the original KDES [1] and 73.81 % for our method. However, both KDES and our method gain nearly the same figures from rank-13 and backward. For iLIDS dataset, the gap between rank-1 recognition rates of the proposed method with AHPE [23] or SDALF (Symmetry- Driven Accumulation of Local Features) [24] is approximately 20 %, and about 7 % with the original KDES [1]. This gap is slightly decreased for KDES but significantly reduced for AHPE and SDALEF after rank-15.
Figure 9. The results of proposed method against AHPE [23], SDALF [24] and KDES [1] on (a) CAVIAR4REID dataset and (b) iLIDS dataset. We compare the results of our proposed method with orig- inal KDES [1] and other state-of-the-art approaches on multi- shot datasets of CAVIAR4REID and iLIDs. The experimental settings are kept the same as in [23], with modified version of iLIDS dataset is selected (including only 69 individuals, with at least 4 images for each) and 72 pedestrians for CAVIAR4REID dataset, in which 50 different individuals are captured under both views. Each view has 10 images for each pedestrian. The outperforming results of the proposed method are shown in Figure 9. For CAVIAR4REID dataset, rank-1 recognition rate of AHPE (Asymmetry-based Histogram Plus Epitome) [23] is much lower than our method. It is only about 8 %, compared with 67.76 % of the original KDES [1] and 73.81 % for our method. However, both KDES and our method gain nearly the same figures from rank-13 and backward. For iLIDS dataset, the gap between rank-1 recognition rates of the proposed method with AHPE [23] or SDALF (Symmetry- Driven Accumulation of Local Features) [24] is approximately 20 %, and about 7 % with the original KDES [1]. This gap is slightly decreased for KDES but significantly reduced for AHPE and SDALEF after rank-15.
Figure 10. The comparative results with (a) reported methods in [25] and (b the results are tested on our dataset.
Figure 10. The comparative results with (a) reported methods in [25] and (b the results are tested on our dataset.
Figure 1. Research model Node: Variable, Edge: Link for two variables, Rectangle: First order indicatot consistent with the basic relationship structure, including stimulus, satisfaction/value, and loyalty/retention, when often applied in consumer research [5]. As follows, we discuss the relevant literature and the development of the hypotheses.
Figure 1. Research model Node: Variable, Edge: Link for two variables, Rectangle: First order indicatot consistent with the basic relationship structure, including stimulus, satisfaction/value, and loyalty/retention, when often applied in consumer research [5]. As follows, we discuss the relevant literature and the development of the hypotheses.
This gives rise to a thinking of an important mediating role of advertising value in the realization of loyalty from the initial drivers. We based on a comparison between the original model and a competing model with extra paths for the antecedents directly to lovaltv. Accordingly. the f2 statistic is based on the
This gives rise to a thinking of an important mediating role of advertising value in the realization of loyalty from the initial drivers. We based on a comparison between the original model and a competing model with extra paths for the antecedents directly to lovaltv. Accordingly. the f2 statistic is based on the
Figure 2. Result of the structure model
Figure 2. Result of the structure model
Figure 2. SUMMIT web app workflow.
Figure 2. SUMMIT web app workflow.
Figure 1. SUMMIT system deployment.
Figure 1. SUMMIT system deployment.
Figure 4. SUMMIT mobile app workflow Prior to the game, users can check out different routes and rewards associated with each of the routes. They can then select a route that provides the rewards they desire and suits their constraints in terms of time and distance. This flexibility enables users to customise their gaming experience based on their needs at any particular time. Fig. 4 shows the workflow of the mobile app. When users select a route, the route information will be downloaded onto their phone assuming Internet connection is available. By pre- loading the routes, the issue of unreliable 3G signal is avoided as the route information is now locally stored, hence will always be available to users when en route. During the hike, only GPS signal is required to track users’ position. Since each checkpoint covers an area of 50-metre radius, a short lost of GPS signals will not affect the performance of the app. These approaches give users the Virtual “Always- On” Connectivity impression allowing them to have an undisrupted interaction experience. The problem of draining the battery power is also minimised as the phone is not constantly connected to the network. Synchronisation with the server occurs the next time network connectivity is available and activated by the user when all logged data on the mobile device is uploaded. To help users locate the rewards, a map that shows the locations of the different checkpoints is provided as illustrated in Figure 5(a). Fig. 5(b) shows a rewards selection dialog box.
Figure 4. SUMMIT mobile app workflow Prior to the game, users can check out different routes and rewards associated with each of the routes. They can then select a route that provides the rewards they desire and suits their constraints in terms of time and distance. This flexibility enables users to customise their gaming experience based on their needs at any particular time. Fig. 4 shows the workflow of the mobile app. When users select a route, the route information will be downloaded onto their phone assuming Internet connection is available. By pre- loading the routes, the issue of unreliable 3G signal is avoided as the route information is now locally stored, hence will always be available to users when en route. During the hike, only GPS signal is required to track users’ position. Since each checkpoint covers an area of 50-metre radius, a short lost of GPS signals will not affect the performance of the app. These approaches give users the Virtual “Always- On” Connectivity impression allowing them to have an undisrupted interaction experience. The problem of draining the battery power is also minimised as the phone is not constantly connected to the network. Synchronisation with the server occurs the next time network connectivity is available and activated by the user when all logged data on the mobile device is uploaded. To help users locate the rewards, a map that shows the locations of the different checkpoints is provided as illustrated in Figure 5(a). Fig. 5(b) shows a rewards selection dialog box.
Figure 5. (a) Map with checkpoints, (b) Rewards selection dialog box
Figure 5. (a) Map with checkpoints, (b) Rewards selection dialog box
Figure 7. Average rating comparison between younger and olderusers. (* denotes variables with significant differences, a = 0.05)
Figure 7. Average rating comparison between younger and olderusers. (* denotes variables with significant differences, a = 0.05)
Figure 6. Overall average rating of all 24 participants.
Figure 6. Overall average rating of all 24 participants.
TABLE I. SYMBOL NOTATION
TABLE I. SYMBOL NOTATION
Figure 3. Star graph example: Hotel Negresco and recommendations on its historical surrounding.
Figure 3. Star graph example: Hotel Negresco and recommendations on its historical surrounding.
Figure 2. Hotel Negresco: information extraction from different sources Figure 1. Sample semantic network: historical relations built around Hotel Negresco.
Figure 2. Hotel Negresco: information extraction from different sources Figure 1. Sample semantic network: historical relations built around Hotel Negresco.
Figure 4. Multi-agent architecture: historical data from various sources and other information are semantically related and analyzed in the smart space. Popular architectural styles for e-Tourism recommender systems are web-based, agent-based, and mobile [1][2]. Smart
Figure 4. Multi-agent architecture: historical data from various sources and other information are semantically related and analyzed in the smart space. Popular architectural styles for e-Tourism recommender systems are web-based, agent-based, and mobile [1][2]. Smart
TABLE II. EXAMPLE OF CONTENT GENERATION to resemble the process of a human using the Web to find other information relevant to these POIs. Finally, augmented reality is used in combination with facet selection to present this POI-related information to an active tourist on her/his mobile device. Similarly to our scenario, this application aims at dy- namic provision of semantically-enriched information in favor of a classical travel guide. In contrast, our scenario introduces both nearby and faraway POIs, which are semantically related within a variety of historical objects, including common histor- ical persons and events. Our scenario supports automation of semantic crawling with POI ranking; the ranks are then used for visual representation of POI recommendations to the user.
TABLE II. EXAMPLE OF CONTENT GENERATION to resemble the process of a human using the Web to find other information relevant to these POIs. Finally, augmented reality is used in combination with facet selection to present this POI-related information to an active tourist on her/his mobile device. Similarly to our scenario, this application aims at dy- namic provision of semantically-enriched information in favor of a classical travel guide. In contrast, our scenario introduces both nearby and faraway POIs, which are semantically related within a variety of historical objects, including common histor- ical persons and events. Our scenario supports automation of semantic crawling with POI ranking; the ranks are then used for visual representation of POI recommendations to the user.
Figure 5. Many KPs interact in a smart space to construct a service.
Figure 5. Many KPs interact in a smart space to construct a service.
Figure 1. Middleware taxonomy, based on [3]. The possibilities of mobile applications increase with the advent of faster hardware, bigger screens and more stable broadband connections. This ongoing evolution of mobile computing leads to larger applications and increases the need for methods reducing software complexity [1]. While reducing complexity has played a central role in software engineering right from the beginning, there is still no silver bullet; com- plexity can only be coped with abstraction [2]. Several types of middleware can aid software developers by providing some of the abstractions that are needed to create nowadays programs.
Figure 1. Middleware taxonomy, based on [3]. The possibilities of mobile applications increase with the advent of faster hardware, bigger screens and more stable broadband connections. This ongoing evolution of mobile computing leads to larger applications and increases the need for methods reducing software complexity [1]. While reducing complexity has played a central role in software engineering right from the beginning, there is still no silver bullet; com- plexity can only be coped with abstraction [2]. Several types of middleware can aid software developers by providing some of the abstractions that are needed to create nowadays programs.
TABLE I. FEATURE OVERVIEW. Legend: +: supported, o: partly supported, -: not supported.
TABLE I. FEATURE OVERVIEW. Legend: +: supported, o: partly supported, -: not supported.
Figure 3. Control flow during application startup and activity change.
Figure 3. Control flow during application startup and activity change.
Figure 4. Interaction on application startup.
Figure 4. Interaction on application startup.
Figure 5. Universal service handles binding of client services. The binding process is shown in Figure 5. In step (1), a binding is requested by a client fragment. The request, containing the classname of the targeted service, is han- dled by the universal service. In step (2), the service is instantiated and its life cycle is executed until onBind() is called, returning a binder object in step (3). The binder object is passed to the corresponding client fragment by calling onServiceConnected() in step (4), finally establishing the connection between client fragment and client service.
Figure 5. Universal service handles binding of client services. The binding process is shown in Figure 5. In step (1), a binding is requested by a client fragment. The request, containing the classname of the targeted service, is han- dled by the universal service. In step (2), the service is instantiated and its life cycle is executed until onBind() is called, returning a binder object in step (3). The binder object is passed to the corresponding client fragment by calling onServiceConnected() in step (4), finally establishing the connection between client fragment and client service.
Figure 6. Overview of the middleware integration architecture. Figure 6 shows an overview of the architecture containing the elements described above. Green dots represent connection endpoints, embodied on Android by implementations of the classes ServiceConnection and Binder.
Figure 6. Overview of the middleware integration architecture. Figure 6 shows an overview of the architecture containing the elements described above. Green dots represent connection endpoints, embodied on Android by implementations of the classes ServiceConnection and Binder.
Figure 1. Path from point 0g to point 01. 0g — 03 — 00 — 01 indicates that the pattern is started from og, then moved to 03, then to 09 and then finally ended at 01.
Figure 1. Path from point 0g to point 01. 0g — 03 — 00 — 01 indicates that the pattern is started from og, then moved to 03, then to 09 and then finally ended at 01.
Figure 3. Perceived strength of patterns by users. It is mostly consistent with computed strength class for patterns.
Figure 3. Perceived strength of patterns by users. It is mostly consistent with computed strength class for patterns.
Figure 2. Choice of different proposed patterns (based on pattern strengths) We used a within-subjects design and asked (“How strong is the following pattern?”’) the participants to rate the strength of six patterns as shown in Figure 2 on a 5-point Likert scale. We chose patterns from three different security levels based on their strength, weak, medium, and strong, as is defined in section IIJ-D. Patterns 2(a) and 2(b) are weak patterns, patterns 2(c) and 2(d) are of medium security level, and patterns 2(e) and 2(f) are strong patterns (refer to Figure 2). To avoid biases, we randomized the order that the patterns were shown to the participants.
Figure 2. Choice of different proposed patterns (based on pattern strengths) We used a within-subjects design and asked (“How strong is the following pattern?”’) the participants to rate the strength of six patterns as shown in Figure 2 on a 5-point Likert scale. We chose patterns from three different security levels based on their strength, weak, medium, and strong, as is defined in section IIJ-D. Patterns 2(a) and 2(b) are weak patterns, patterns 2(c) and 2(d) are of medium security level, and patterns 2(e) and 2(f) are strong patterns (refer to Figure 2). To avoid biases, we randomized the order that the patterns were shown to the participants.
TABLE I. GROUP DEMOGRAPHICS. We recruited a total of 270 US-based workers to participate in our experiment. Note, that these are different participants than the participants described in Section III-A. We also con- fined our user study to those who are familiar with the Android Patterns. Demographics of the participants and number of participants are given in Table I. The user study procedure involved two main steps a) Registration - Participants were assigned randomly to one of the groups, i.e, NORMAL, EPSM or BLINK. After that, they were asked to choose a pattern and then to verify it immediately. All participants were instructed to imagine that they have received a new phone and would like to set an Android Pattern on it. They were asked to choose a pattern of a minimum length four. b) Survey - After creating a pattern, participants were asked to complete a survey. We paid $0.40 to each participant upon the completion of our user study.
TABLE I. GROUP DEMOGRAPHICS. We recruited a total of 270 US-based workers to participate in our experiment. Note, that these are different participants than the participants described in Section III-A. We also con- fined our user study to those who are familiar with the Android Patterns. Demographics of the participants and number of participants are given in Table I. The user study procedure involved two main steps a) Registration - Participants were assigned randomly to one of the groups, i.e, NORMAL, EPSM or BLINK. After that, they were asked to choose a pattern and then to verify it immediately. All participants were instructed to imagine that they have received a new phone and would like to set an Android Pattern on it. They were asked to choose a pattern of a minimum length four. b) Survey - After creating a pattern, participants were asked to complete a survey. We paid $0.40 to each participant upon the completion of our user study.
Figure 4. Sequence showing the change in colors for a single pattern, as user draw it strength of the pattern, the color of the pattern gets updated (see Section II-D for details on pattern strength). Note that change in the color usually goes from red to yellow to green, however in certain cases it can sometimes goes the other way around, i.e., from green to red. For example, a half-drawn Z (09 — 01 — 02 — 04 — 0g — 07) iS a More secure pattern than a full drawn Z (09 — 01 — 02 — 04 — 06 — O7 — Og).
Figure 4. Sequence showing the change in colors for a single pattern, as user draw it strength of the pattern, the color of the pattern gets updated (see Section II-D for details on pattern strength). Note that change in the color usually goes from red to yellow to green, however in certain cases it can sometimes goes the other way around, i.e., from green to red. For example, a half-drawn Z (09 — 01 — 02 — 04 — 0g — 07) iS a More secure pattern than a full drawn Z (09 — 01 — 02 — 04 — 06 — O7 — Og).
Figure 5. Suggesting a start point by blinking In a pilot study, we examined a number of techniques to suggest a starting point to the users without hampering the usability of Android Patterns. Based on our observations and users suggestions, we concluded that a) suggestion by blinking a point is very effective, and b) users need to be told what is expected be done with the suggested point without hampering the usability of the system. Our final design uses a blinking point (see Figure 5) and similar to EPSM it pops-up a recommendation message stating “Jt is recommended to start your pattern with the blinking point but NOT MANDATORY”, as it helps user to understand what to do and how to proceed with the blinking point.
Figure 5. Suggesting a start point by blinking In a pilot study, we examined a number of techniques to suggest a starting point to the users without hampering the usability of Android Patterns. Based on our observations and users suggestions, we concluded that a) suggestion by blinking a point is very effective, and b) users need to be told what is expected be done with the suggested point without hampering the usability of the system. Our final design uses a blinking point (see Figure 5) and similar to EPSM it pops-up a recommendation message stating “Jt is recommended to start your pattern with the blinking point but NOT MANDATORY”, as it helps user to understand what to do and how to proceed with the blinking point.
Figure 7. Strength level of the patterns created by users of NORMAL, BLINK and EPSM. Red, yellow, and green bars indicate weak, moderate, and strong patterns respectively.
Figure 7. Strength level of the patterns created by users of NORMAL, BLINK and EPSM. Red, yellow, and green bars indicate weak, moderate, and strong patterns respectively.
Figure 6. Distribution of starting point of the patterns in the NORMAL, BLINK and EPSM. BLINK removes the bias of starting points. Figure 6(a) shows the percentage of patterns starting from each of the nine points in all three schemes. As expected in NORMAL and EPSM, we can observe that the starting point probabilities of the top two corner points (52.1% and 35.7%) are much higher than the other points. This is because there was no recommendation provided for eliminating the starting point bias in NORMAL and EPSM.
Figure 6. Distribution of starting point of the patterns in the NORMAL, BLINK and EPSM. BLINK removes the bias of starting points. Figure 6(a) shows the percentage of patterns starting from each of the nine points in all three schemes. As expected in NORMAL and EPSM, we can observe that the starting point probabilities of the top two corner points (52.1% and 35.7%) are much higher than the other points. This is because there was no recommendation provided for eliminating the starting point bias in NORMAL and EPSM.
Figure 9. Recall accuracy (in terms of trials before success)
Figure 9. Recall accuracy (in terms of trials before success)
Figure 8. Guessing attack against NORMAL, BLINK and EPSM. Guessing model is built over the NORMAL version. needs 3344 and 1918 guesses to guess 50% of the patterns in EPSM and BLINK, respectively. In terms of partial entropy of patterns, this translates to 9.79 bits of entropy for the patterns used by 50% of NORMAL, and 11.7 and 10.9 bits of entropy for the same proportion of users in EPSM and BLINK.
Figure 8. Guessing attack against NORMAL, BLINK and EPSM. Guessing model is built over the NORMAL version. needs 3344 and 1918 guesses to guess 50% of the patterns in EPSM and BLINK, respectively. In terms of partial entropy of patterns, this translates to 9.79 bits of entropy for the patterns used by 50% of NORMAL, and 11.7 and 10.9 bits of entropy for the same proportion of users in EPSM and BLINK.
Figure 1. A typical RPL DODAG with one root and six other nodes.
Figure 1. A typical RPL DODAG with one root and six other nodes.
ABLE II. MLPLW MEMORY CONSUMPTION TABLE
ABLE II. MLPLW MEMORY CONSUMPTION TABLE
Figure 2. System flowchart After calculating the position of the object, the drone will compute a new position for it to fly to (if the object has moved) so that it will continue to have the object in its view. This process can be repeated as often as necessary. Figure 2 gives a flowchart of the system. B. Object detection
Figure 2. System flowchart After calculating the position of the object, the drone will compute a new position for it to fly to (if the object has moved) so that it will continue to have the object in its view. This process can be repeated as often as necessary. Figure 2 gives a flowchart of the system. B. Object detection
Our system is not restricted to work on only one class of objects. It can be extended to work on any class of objects, but we need to have a reliable object detection algorithm for it. In this paper, we demonstrate our system on pedestrians, and we make use of the Histogram of Oriented Gradients (HOG) pedestrian descriptor from the OpenCV 2.4.10 library (“peopledetect.cpp”). Figure 3 shows an example of the pedestrian detection algorithm in OpenCV, which displays a bounding box over the detected pedestrian. Figure 3. Example of pedestrian detection from the drone
Our system is not restricted to work on only one class of objects. It can be extended to work on any class of objects, but we need to have a reliable object detection algorithm for it. In this paper, we demonstrate our system on pedestrians, and we make use of the Histogram of Oriented Gradients (HOG) pedestrian descriptor from the OpenCV 2.4.10 library (“peopledetect.cpp”). Figure 3 shows an example of the pedestrian detection algorithm in OpenCV, which displays a bounding box over the detected pedestrian. Figure 3. Example of pedestrian detection from the drone
Figure 1 shows an overview of our system. Our system consists of a drone mounted with a single camera at position (xa, ya). Using the algorithm described later in this section, the drone will compute the distance Z and angle 6,of an object (¢.g., pedestrian) relative to itself, and then use this distance to compute the position of the object (x, y) in global coordinates. We assume that the position of the drone is always known (e.g., through GPS, or WiFi positioning, or from Inertial Navigation System (INS), if available).
Figure 1 shows an overview of our system. Our system consists of a drone mounted with a single camera at position (xa, ya). Using the algorithm described later in this section, the drone will compute the distance Z and angle 6,of an object (¢.g., pedestrian) relative to itself, and then use this distance to compute the position of the object (x, y) in global coordinates. We assume that the position of the drone is always known (e.g., through GPS, or WiFi positioning, or from Inertial Navigation System (INS), if available).
Table I shows the data measured from Figure 6. While the Op ratio is not constant throughout the entire image, the variation was found to be quite small (0.001°/pixel) between the two cases where the object is in the image center and at the image edge.
Table I shows the data measured from Figure 6. While the Op ratio is not constant throughout the entire image, the variation was found to be quite small (0.001°/pixel) between the two cases where the object is in the image center and at the image edge.
Figure 4. Classical disparity formula diagram In order to obtain an accurate estimation of the depth of an object, the camera needs to be calibrated. The basic concept of the classical structure-from-motion algorithm is from the geometry shown in Figure 4.
Figure 4. Classical disparity formula diagram In order to obtain an accurate estimation of the depth of an object, the camera needs to be calibrated. The basic concept of the classical structure-from-motion algorithm is from the geometry shown in Figure 4.
Figure 5. Geometry of view angle calibration
Figure 5. Geometry of view angle calibration
Figure 6. Images of an object taken at various positions We can obtain this value by placing an object at various positions in the drone’s field of view (Figure 6) and measuring the image offset p and the angle subtended @
Figure 6. Images of an object taken at various positions We can obtain this value by placing an object at various positions in the drone’s field of view (Figure 6) and measuring the image offset p and the angle subtended @
Figure 7. Classical stereo vision method for the stationary pedestrian We will first instruct the drone to takeoff, and then hover for a few seconds for it to stabilize itself. Then, we will call the object detector function to find any objects within its view. For the ease of discussion, let us use the example of pedestrian detection. After the drone is stabilized, we call the HOG pedestrian detector from the OpenCV library. If more than one pedestrian is found, we need to choose which pedestrian is the one that we are trying to localize. After we have found the pedestrian, we rotate the drone (yaw) so that the pedestrian (i.e., the centroid of the bounding box) is in the center of the image (see Figure 2 for an overview of the system).
Figure 7. Classical stereo vision method for the stationary pedestrian We will first instruct the drone to takeoff, and then hover for a few seconds for it to stabilize itself. Then, we will call the object detector function to find any objects within its view. For the ease of discussion, let us use the example of pedestrian detection. After the drone is stabilized, we call the HOG pedestrian detector from the OpenCV library. If more than one pedestrian is found, we need to choose which pedestrian is the one that we are trying to localize. After we have found the pedestrian, we rotate the drone (yaw) so that the pedestrian (i.e., the centroid of the bounding box) is in the center of the image (see Figure 2 for an overview of the system).
The baseline between any two images is simply the sum of the values between them. For example, the baseline between image 0 and image 5 is 47.50451.75+75.75491.25 = 266.25(cm). ER SIO MEETS errr EI EE IEEE ERE ES We repeated ‘the experiment three times, and averaged the results of the four experiments to produce the baselines shown in Table II. To measure the length of the baselines, we placed numerous markers on a wall, and made the drone fly parallel to the wall, taking images as mentioned above. From the images, we can determine the positions of the drone where the images were taken. We observe that the first image (i.e., at position 0) was exactly the same as the image at hovering (i.e., it was taken before the drone has even started to rotate (roll)), and the second image (i.e., at position 1) was taken just after the drone has completed its rotation (roll) to the left (but before any horizontal translation takes place — so no change in its position). The rest of the images (at positions 2 to 10) were of increasing distances between each other, which is due to the inertial and the acceleration of the drone, which needed some time to accelerate to the desired speed.
The baseline between any two images is simply the sum of the values between them. For example, the baseline between image 0 and image 5 is 47.50451.75+75.75491.25 = 266.25(cm). ER SIO MEETS errr EI EE IEEE ERE ES We repeated ‘the experiment three times, and averaged the results of the four experiments to produce the baselines shown in Table II. To measure the length of the baselines, we placed numerous markers on a wall, and made the drone fly parallel to the wall, taking images as mentioned above. From the images, we can determine the positions of the drone where the images were taken. We observe that the first image (i.e., at position 0) was exactly the same as the image at hovering (i.e., it was taken before the drone has even started to rotate (roll)), and the second image (i.e., at position 1) was taken just after the drone has completed its rotation (roll) to the left (but before any horizontal translation takes place — so no change in its position). The rest of the images (at positions 2 to 10) were of increasing distances between each other, which is due to the inertial and the acceleration of the drone, which needed some time to accelerate to the desired speed.
After we have obtained the depth Z and angle 6 we can use them together with the drone position to calculate the global coordinates of the pedestrian. Figure 10 shows the geometry for calculating the coordinate transformation for each of the 3 cases. In Figure 10, (%qg2,Va2,@q) and (%a1,Va1,9q) refers the positions of the drone (which is assumed to be known). (Xs,¥s), (Xp, Mp), (Xv, Vy) are the positions of the pedestrian in each of the three cases of stationary, moving parallel, or moving perpendicular (vertical) to the baseline, respectively Note that the initial coordinate of the drone (x41, Yai, 9a) is vital which is used as boundary condition for coordinate transformation. From this geometry, we can directly
After we have obtained the depth Z and angle 6 we can use them together with the drone position to calculate the global coordinates of the pedestrian. Figure 10 shows the geometry for calculating the coordinate transformation for each of the 3 cases. In Figure 10, (%qg2,Va2,@q) and (%a1,Va1,9q) refers the positions of the drone (which is assumed to be known). (Xs,¥s), (Xp, Mp), (Xv, Vy) are the positions of the pedestrian in each of the three cases of stationary, moving parallel, or moving perpendicular (vertical) to the baseline, respectively Note that the initial coordinate of the drone (x41, Yai, 9a) is vital which is used as boundary condition for coordinate transformation. From this geometry, we can directly
Figure 9. Geometry for pedestrian moving perpendicular to the baseline
Figure 9. Geometry for pedestrian moving perpendicular to the baseline
Figure 8. Geometry for pedestrian moving parallel to the baseline The key idea here is that we now need three images, at position 0, 1 and n (the last image where the pedestrian can still be seen) to compute the depth Z. Due to the fact that at position 1 the drone has just completed its rotation (roll) to the left, the drone has not moved yet but the pedestrian had moved by a distance C. This gives us the very important image offset p: that allows us to calculate @. At the nth position, at angle &, we know that the pedestrian has moved by an additional distance of Cx (n—1). So from the geometry, we can generate the following relation:
Figure 8. Geometry for pedestrian moving parallel to the baseline The key idea here is that we now need three images, at position 0, 1 and n (the last image where the pedestrian can still be seen) to compute the depth Z. Due to the fact that at position 1 the drone has just completed its rotation (roll) to the left, the drone has not moved yet but the pedestrian had moved by a distance C. This gives us the very important image offset p: that allows us to calculate @. At the nth position, at angle &, we know that the pedestrian has moved by an additional distance of Cx (n—1). So from the geometry, we can generate the following relation:
TABLE HI. STATIONARY PEDESTRIAN “ABLE IV. PERPENDICULARLY MOVING PEDESTRIAN
TABLE HI. STATIONARY PEDESTRIAN “ABLE IV. PERPENDICULARLY MOVING PEDESTRIAN
Figure 12. Examples of Image Sequence and pedestrian detection (5.5m, stationary pedestrian) stationary) are shown in Figure 12. Notice that between the first and second image, there is only a rotation of the drone and there is no horizontal movement.
Figure 12. Examples of Image Sequence and pedestrian detection (5.5m, stationary pedestrian) stationary) are shown in Figure 12. Notice that between the first and second image, there is only a rotation of the drone and there is no horizontal movement.
The drone used in this study is the Parrot AR.Drone2 GPS edition [10]. It has a forward-looking 720p HD camera and a vertical QVGA camera. The drone is controlled from an Intel i5 laptop running Windows 8.1 with 4GB of RAM. The software we used to control the drone is the “CV Drone” package which is available from Github [11]. The image processing routines were from the OpenCV 2.4.10 library, and the entire system was developed in Microsoft Visual Studio 2013.
The drone used in this study is the Parrot AR.Drone2 GPS edition [10]. It has a forward-looking 720p HD camera and a vertical QVGA camera. The drone is controlled from an Intel i5 laptop running Windows 8.1 with 4GB of RAM. The software we used to control the drone is the “CV Drone” package which is available from Github [11]. The image processing routines were from the OpenCV 2.4.10 library, and the entire system was developed in Microsoft Visual Studio 2013.
calculate the position of a pedestrian’s position using one of! the three following equations.
calculate the position of a pedestrian’s position using one of! the three following equations.
In the second scenario, we have two or more images. The first image may be generated by a person with a digital camera or mobile phone camera, or by another intelligent CCTV camera (CCTV1). The second image may be taken by a drone (i.e., UAV) or by another CCTV camera. If the drone is the source of the second image, it can offer not only the top view but also many views in different angles because it is a flying camera. These images are then sent to CCTV2 where it will generate a new view of the car that matches its own viewpoint of the road. In the first scenario, we have only one image of the car. The image may be generated by a person with a digital camera or mobile phone camera, or even by another intelligent CCTV camera (CCTV1). This image is then sent to CCTV2 where it will generate a new view of the car that matches its own viewpoint of the road.
In the second scenario, we have two or more images. The first image may be generated by a person with a digital camera or mobile phone camera, or by another intelligent CCTV camera (CCTV1). The second image may be taken by a drone (i.e., UAV) or by another CCTV camera. If the drone is the source of the second image, it can offer not only the top view but also many views in different angles because it is a flying camera. These images are then sent to CCTV2 where it will generate a new view of the car that matches its own viewpoint of the road. In the first scenario, we have only one image of the car. The image may be generated by a person with a digital camera or mobile phone camera, or even by another intelligent CCTV camera (CCTV1). This image is then sent to CCTV2 where it will generate a new view of the car that matches its own viewpoint of the road.
We identify that CCTV1 must provide at least one image of the car (with no restriction to its viewpoint) to CCTV2. Since CCTV2 knows its own position, elevation and camera inclination, it must generate a view of the car that matches its viewpoint of the road when the car begins to come into its field of view. Figures 2 and 3 illustrate two scenarios where the source image(s) can be obtained.
We identify that CCTV1 must provide at least one image of the car (with no restriction to its viewpoint) to CCTV2. Since CCTV2 knows its own position, elevation and camera inclination, it must generate a view of the car that matches its viewpoint of the road when the car begins to come into its field of view. Figures 2 and 3 illustrate two scenarios where the source image(s) can be obtained.
Figure 4. System overview
Figure 4. System overview
Figure 7 shows another Opel Corsa that has a different pattern from Figure 6. This Opel Corsa has a black plate with a green stripe mark running from the hood to the trunk. In this case, there are also two images of the same car. The second image (Figure 7(b)) represents an image taken by a drone or from a CCTV camera that is mounted at the top of a building. If this image was not available, we will face the same problem as the Opel Corsa with Flame patterns. The white points in Figure 7(a) and (b) are the points of the polygon where the user has to identify and enter into the system. If both images (Figure 6(a) and (b)) are available then we can completely reconstruct the 3D model of the Opel Corsa. The white circles in Figure 6 are the points of the polygon where the user has to identify and enter into the system.
Figure 7 shows another Opel Corsa that has a different pattern from Figure 6. This Opel Corsa has a black plate with a green stripe mark running from the hood to the trunk. In this case, there are also two images of the same car. The second image (Figure 7(b)) represents an image taken by a drone or from a CCTV camera that is mounted at the top of a building. If this image was not available, we will face the same problem as the Opel Corsa with Flame patterns. The white points in Figure 7(a) and (b) are the points of the polygon where the user has to identify and enter into the system. If both images (Figure 6(a) and (b)) are available then we can completely reconstruct the 3D model of the Opel Corsa. The white circles in Figure 6 are the points of the polygon where the user has to identify and enter into the system.
the hood and driver-side door patterns and we will not know that there are also patterns on the trunk and the passenger-side door. This may lead to an incorrect 3D model reconstruction. To increase our chances for obtaining a correct 3D model, we can make certain assumptions (e.g., that the flame mark exists on the doors on both side of the car) and generate several candidate models for recognition.
the hood and driver-side door patterns and we will not know that there are also patterns on the trunk and the passenger-side door. This may lead to an incorrect 3D model reconstruction. To increase our chances for obtaining a correct 3D model, we can make certain assumptions (e.g., that the flame mark exists on the doors on both side of the car) and generate several candidate models for recognition.
Figure 5. Audi Q7 with strips on the hood Figure 5 shows an Audi Q7 that has stripes running down the hood. The four points in red (on the hood near to the stripes) in Figure 5 are the four points that the human user needs to enter into the system.
Figure 5. Audi Q7 with strips on the hood Figure 5 shows an Audi Q7 that has stripes running down the hood. The four points in red (on the hood near to the stripes) in Figure 5 are the four points that the human user needs to enter into the system.
Figure 10. Opel Corsa with warped image of green stripes
Figure 10. Opel Corsa with warped image of green stripes
To perform the texture mapping, we need to attach the warped image to the UVW map of the car. Here, the human user has to enter another 4 points on the UVW map to show where the warped image will go. Figure 11(b) shows the 4 points that the human user have entered. After we have downloaded the corresponding car model, we need to perform a UVW unwrap operation (from any 3D modeling software such as 3DSmax) to obtain the UVW map of the model. The UVW map allows us to perform texture mapping onto the 3D car model. An example of the UVW map of the Audi Q7 is showed in Figure 11(a).
To perform the texture mapping, we need to attach the warped image to the UVW map of the car. Here, the human user has to enter another 4 points on the UVW map to show where the warped image will go. Figure 11(b) shows the 4 points that the human user have entered. After we have downloaded the corresponding car model, we need to perform a UVW unwrap operation (from any 3D modeling software such as 3DSmax) to obtain the UVW map of the model. The UVW map allows us to perform texture mapping onto the 3D car model. An example of the UVW map of the Audi Q7 is showed in Figure 11(a).
Figure 9. Opel Corsa with warped image of flame pattern Figure 8(a) is original car image of Audi Q7 which is not modified at all. The Figure 8(b) is the warped image of extracted black stripe mark from Figure 8(a).
Figure 9. Opel Corsa with warped image of flame pattern Figure 8(a) is original car image of Audi Q7 which is not modified at all. The Figure 8(b) is the warped image of extracted black stripe mark from Figure 8(a).
Figure 12. Completed UVW map of the Audi Q7 ATter entering the 4 points, the warped Image Wl! be automatically resized and then “pasted” onto the UVW map. In addition, the human user needs to specify one more point in the image (that has the color of the body of the car) so that all the panels in the UVW can be filled with this color. This completes the texture mapping process on the UVW map and we are ready to reconstruct the 3D car. Figures 12, 13 and 14 show the completed UVW maps of all the three cars in our experiment.
Figure 12. Completed UVW map of the Audi Q7 ATter entering the 4 points, the warped Image Wl! be automatically resized and then “pasted” onto the UVW map. In addition, the human user needs to specify one more point in the image (that has the color of the body of the car) so that all the panels in the UVW can be filled with this color. This completes the texture mapping process on the UVW map and we are ready to reconstruct the 3D car. Figures 12, 13 and 14 show the completed UVW maps of all the three cars in our experiment.
With the information on the car position and the camera position, we can generate the expected view of the car from CCTV 2 using the gluLookat function in the OpenGL library. By rotating the 3D reconstructed model, we can generate many arbitrary views of the car at different angles.
With the information on the car position and the camera position, we can generate the expected view of the car from CCTV 2 using the gluLookat function in the OpenGL library. By rotating the 3D reconstructed model, we can generate many arbitrary views of the car at different angles.
Figure 16 shows the results of reconstructing the 3D model of the three cars to the same view as the original images so that we can verify that the result is correct.
Figure 16 shows the results of reconstructing the 3D model of the three cars to the same view as the original images so that we can verify that the result is correct.
Figure 14. Completed UVW map of the Opel Corsa with Green Stripes Figures 13(a), 13(b) and 13(c) are completed UVW map of Opel Corsa with flame pattern. Each of them is generated by attaching warped images of part D to the region of UVW which is chosen by user. Figure 9(b) is used for Figure 14(a), Figures 9(c) and 9(e) are used for Figure 14(b), and Figure 9(f) is used for Figure 13(c). Figures 14(a), 14(b) and 14(c) are the completed UVW map of the Opel Corsa with green stripes. The resized Figures 10(b), 10(d), and 10(e) are attached to corresponding position of UVW map of Opel Corsa.
Figure 14. Completed UVW map of the Opel Corsa with Green Stripes Figures 13(a), 13(b) and 13(c) are completed UVW map of Opel Corsa with flame pattern. Each of them is generated by attaching warped images of part D to the region of UVW which is chosen by user. Figure 9(b) is used for Figure 14(a), Figures 9(c) and 9(e) are used for Figure 14(b), and Figure 9(f) is used for Figure 13(c). Figures 14(a), 14(b) and 14(c) are the completed UVW map of the Opel Corsa with green stripes. The resized Figures 10(b), 10(d), and 10(e) are attached to corresponding position of UVW map of Opel Corsa.
Figure 13. Completed UVW map of the Opel Corsa with Flame pattern each other, the algorithm performs resizing process before pasting.
Figure 13. Completed UVW map of the Opel Corsa with Flame pattern each other, the algorithm performs resizing process before pasting.
Figure 18. Merging with CCTV background image We can see that the car images blend in very well to the CCTV background image and the resulting image looks quite realistic. We also showed a zoomed-in image of the Opel Corsa with green stripes to show that the
Figure 18. Merging with CCTV background image We can see that the car images blend in very well to the CCTV background image and the resulting image looks quite realistic. We also showed a zoomed-in image of the Opel Corsa with green stripes to show that the
Figures 16(a), 16(c), 16(e), 16(g) and 16(i) are the raw car images of Audi Q7 and Opel Corsa which are same with Figures 8(a), 9(d), 9(a), 10(a) and 10(c) each. Figures 16(b), 16(d), 16(f), 16(h) and 16(j) are the reconstructed 3D model of each car.
Figures 16(a), 16(c), 16(e), 16(g) and 16(i) are the raw car images of Audi Q7 and Opel Corsa which are same with Figures 8(a), 9(d), 9(a), 10(a) and 10(c) each. Figures 16(b), 16(d), 16(f), 16(h) and 16(j) are the reconstructed 3D model of each car.
Figure 17 shows the results of generating new arbitrary views of the three cars. Figures 17(a) and 17(b) are the arbitrary view of Audi Q7. Figures 17(c), 17(d), 17(e) and 17(f) are different view of Opel Corsa with flame pattern and green stripe. We can see that all the three cars in our experiment look realistic.
Figure 17 shows the results of generating new arbitrary views of the three cars. Figures 17(a) and 17(b) are the arbitrary view of Audi Q7. Figures 17(c), 17(d), 17(e) and 17(f) are different view of Opel Corsa with flame pattern and green stripe. We can see that all the three cars in our experiment look realistic.
Figure 1. A conceptual view of a Disaster Warning System The main goal of a disaster warning system is preventing loss of properties by providing suitable information to people who can be affected by disaster events [2][7]. To achieve this goal, the system should analyze the impact of disaster events in detail, pick up all relevant people to the disaster, prepare proper messages and send the messages reliably in timely manner.
Figure 1. A conceptual view of a Disaster Warning System The main goal of a disaster warning system is preventing loss of properties by providing suitable information to people who can be affected by disaster events [2][7]. To achieve this goal, the system should analyze the impact of disaster events in detail, pick up all relevant people to the disaster, prepare proper messages and send the messages reliably in timely manner.
Figure 2. The architecture of the proposed disaster warnding system disaster warning system. Finally, the failure-resilient geo- aware dissemination module is related to the dissemination component of a disaster warning system.
Figure 2. The architecture of the proposed disaster warnding system disaster warning system. Finally, the failure-resilient geo- aware dissemination module is related to the dissemination component of a disaster warning system.
Figure 4. Propagation process with different channel preferences When the message propagation is considered, we can consider the multi-modal channels for the message propagation, because users usually use various channels to propagate messages through social connections [6]. In an emergency situation, the message propagation may enhance the possibility that a message reaches to the target user within a given time limit.
Figure 4. Propagation process with different channel preferences When the message propagation is considered, we can consider the multi-modal channels for the message propagation, because users usually use various channels to propagate messages through social connections [6]. In an emergency situation, the message propagation may enhance the possibility that a message reaches to the target user within a given time limit.
Figure 3. Performance of predicting user attitute (E: Extraverstion and I: Introversion) in MBTI (Myer-Briggs Type Indicator). SVM (Support Vector Machine) is used as a classifier and each location clustering technique is used to extract classification features.
Figure 3. Performance of predicting user attitute (E: Extraverstion and I: Introversion) in MBTI (Myer-Briggs Type Indicator). SVM (Support Vector Machine) is used as a classifier and each location clustering technique is used to extract classification features.
Figure 5. Reliability of message dissemination under large failures In a disaster event, there is a high probability that the network devices are ruined. In this situation, how to provide a reliable sending method is also important. For example, a mobile phone is a reliable method to send the message to user in the normal case. But in a disaster, the mobile phone can be ruined or the user cannot get the message. In this case, we need to choose another way to send the message to the user. Also, it is possible that the network devices such as switches or routers are destroyed. To send message to the user, we might have to reroute the message to another router.
Figure 5. Reliability of message dissemination under large failures In a disaster event, there is a high probability that the network devices are ruined. In this situation, how to provide a reliable sending method is also important. For example, a mobile phone is a reliable method to send the message to user in the normal case. But in a disaster, the mobile phone can be ruined or the user cannot get the message. In this case, we need to choose another way to send the message to the user. Also, it is possible that the network devices such as switches or routers are destroyed. To send message to the user, we might have to reroute the message to another router.
TABLE I. NOTATIONS AND SHORTHANDS REFERENCE. where X®SS!(q) in Decibel (dB) is the Received Signal Strength Indicator (RSS) at distance d, dp is the reference distance, 7 is the path loss exponent (PLE), and «, ~ (0,07) is a random variable reflecting the attenuation caused by flat fading [14].
TABLE I. NOTATIONS AND SHORTHANDS REFERENCE. where X®SS!(q) in Decibel (dB) is the Received Signal Strength Indicator (RSS) at distance d, dp is the reference distance, 7 is the path loss exponent (PLE), and «, ~ (0,07) is a random variable reflecting the attenuation caused by flat fading [14].
the location misclassification rate weighted by the priority and visiting frequency is minimized. Intuitively, the more fingerprints a) collected the better we can estimate the mean at that lovation Aj. and the less the misclassification rate (refer Section II.B for the relations). By the maximum likelihood method, our estimates of Ay. is the sample mean of fingerprints ao), which is normally distributed with mean Aj. and covariance 1 Mis where n,; is the number of fingerprints collected. By Assumption 1, %, is the same for all locations i € [N]. Consider m and I to be neighboring two TeBIONS, then by applying the results in (7) and (9) with U,, = a and yy) = oH d (dis diagonal with entries c? by Assumption 1),
the location misclassification rate weighted by the priority and visiting frequency is minimized. Intuitively, the more fingerprints a) collected the better we can estimate the mean at that lovation Aj. and the less the misclassification rate (refer Section II.B for the relations). By the maximum likelihood method, our estimates of Ay. is the sample mean of fingerprints ao), which is normally distributed with mean Aj. and covariance 1 Mis where n,; is the number of fingerprints collected. By Assumption 1, %, is the same for all locations i € [N]. Consider m and I to be neighboring two TeBIONS, then by applying the results in (7) and (9) with U,, = a and yy) = oH d (dis diagonal with entries c? by Assumption 1),
is based on the HBSM (line 5, see Section II) and weighted misclassification cost (line 8, see Section III).
is based on the HBSM (line 5, see Section II) and weighted misclassification cost (line 8, see Section III).
Figure 2. Expected loss of fingerprints allocation 9f? for each router setup (out of 84 candidates) with Nzo+ = 15 total budget, N = 5 subregions, kK = 83 routers, as selected by APEC Exhaust, APEC Greedy (with different batch sizes B), and random approach (black error bar showing means standard deviation over 20 test runs). E]
Figure 2. Expected loss of fingerprints allocation 9f? for each router setup (out of 84 candidates) with Nzo+ = 15 total budget, N = 5 subregions, kK = 83 routers, as selected by APEC Exhaust, APEC Greedy (with different batch sizes B), and random approach (black error bar showing means standard deviation over 20 test runs). E]
Figure 1. Map of the toy example showing the access points locations and fingerprints allocations (radius of the green circle) determined by APEC Exhaust (also APEC Greedy since they agree). We design a simple problem, as shown in Figure 1, with M = 9 possible router locations and N = 5 subregions, where the location priority is color coded, and the frequency is identical for all 5 subregions. The total number of finger- prints N54 is restricted so that APEC Exhaust is tractable. A random approach is also implemented where the fingerprints are distributed randomly.
Figure 1. Map of the toy example showing the access points locations and fingerprints allocations (radius of the green circle) determined by APEC Exhaust (also APEC Greedy since they agree). We design a simple problem, as shown in Figure 1, with M = 9 possible router locations and N = 5 subregions, where the location priority is color coded, and the frequency is identical for all 5 subregions. The total number of finger- prints N54 is restricted so that APEC Exhaust is tractable. A random approach is also implemented where the fingerprints are distributed randomly.
Figure 3. Expected loss vs. router setup indexed by the optimal costs given by solving the fully coupled problem (APEC Exhaust). The plot also shows the loss estimated by random selection (green, mean + | standard deviation) and heuristic router selection (red). Another possible heuristic, random selection, is to treat each setup uniformly, where they have identical distribution of expected loss, as shown in Figure 3. We might end up in a region, i.e., the right half of the graph, where even with the optimized fingerprints of the incurred loss is significantly higher then the optimal, since unlike the decoupling heuristic, the ranking information is lost.
Figure 3. Expected loss vs. router setup indexed by the optimal costs given by solving the fully coupled problem (APEC Exhaust). The plot also shows the loss estimated by random selection (green, mean + | standard deviation) and heuristic router selection (red). Another possible heuristic, random selection, is to treat each setup uniformly, where they have identical distribution of expected loss, as shown in Figure 3. We might end up in a region, i.e., the right half of the graph, where even with the optimized fingerprints of the incurred loss is significantly higher then the optimal, since unlike the decoupling heuristic, the ranking information is lost.
Figure 4. Plots of RSSI vs. log distance color coded by devices. A linear regression line is fitted on the data: RSSI = —8.31 - log(d) — 53.4. Figure 4 shows the RSSI with respect to the log distance from the fingerprint to the access point. Generally, a linear relation is observed (with correlation score .77), though the variance can be reduced by accounting for walls as suggested by Bahl and Padmanabhan, [1]. It can be seen that the dependency is stable for all devices (Android, iPad, iPhone), which ensures cross-platform performance of fingerprints con- figuration.
Figure 4. Plots of RSSI vs. log distance color coded by devices. A linear regression line is fitted on the data: RSSI = —8.31 - log(d) — 53.4. Figure 4 shows the RSSI with respect to the log distance from the fingerprint to the access point. Generally, a linear relation is observed (with correlation score .77), though the variance can be reduced by accounting for walls as suggested by Bahl and Padmanabhan, [1]. It can be seen that the dependency is stable for all devices (Android, iPad, iPhone), which ensures cross-platform performance of fingerprints con- figuration.
Figure 5. Plots of the actual cost (12) vs. the expected cost (11), color coded by the number of routers deployed. A quadratic curve is fitted to the data: L=0.37- Ef — 2.34-L- + 4.46, where L and Le are actual and expected losses given in (12) and (11) respectively.
Figure 5. Plots of the actual cost (12) vs. the expected cost (11), color coded by the number of routers deployed. A quadratic curve is fitted to the data: L=0.37- Ef — 2.34-L- + 4.46, where L and Le are actual and expected losses given in (12) and (11) respectively.
Figure 7. Visualization of IPS configuration in CREST for experiment A (left) and B (right), where 5 or 7 access points are deployed according to APEC Greedy. Location priority map is color coded as illustrated in the legend. Location frequency map is not shown, which is HIGH for cubicles, MED for shared spaces, and LOW for corridors. The radius of the green circle indicates the number of fingerprints at the spot. Figure 6. Actual cost distribution in Experiment A (left) and B (right) for all devices color coded by the fingerprinting methods, i.e., random selection, uniform selection, and APEC Greedy. The box goes from the 1°* quantile to the 3"@ quantile, with the black line indicating the mean. As a side note, Android device incurs less cost in both experiments as a result of its stability of the RSSI signal.
Figure 7. Visualization of IPS configuration in CREST for experiment A (left) and B (right), where 5 or 7 access points are deployed according to APEC Greedy. Location priority map is color coded as illustrated in the legend. Location frequency map is not shown, which is HIGH for cubicles, MED for shared spaces, and LOW for corridors. The radius of the green circle indicates the number of fingerprints at the spot. Figure 6. Actual cost distribution in Experiment A (left) and B (right) for all devices color coded by the fingerprinting methods, i.e., random selection, uniform selection, and APEC Greedy. The box goes from the 1°* quantile to the 3"@ quantile, with the black line indicating the mean. As a side note, Android device incurs less cost in both experiments as a result of its stability of the RSSI signal.
Figure 1: Customization based on semantic technologies [22].
Figure 1: Customization based on semantic technologies [22].
Figure 2: Production scenario. different production lines, which are coupled via a workpiece carrier (WPC). The WPC accompanies the product through the milling, assembling and processing cycle and carries the product physically. The WPC is also equipped with a CPS- enabled ADOMe, that couples the physical product part with its virtual counterpart, which represents all product-specific data. Within this interconnected infrastructure, the WPC has access to all information of the product, to provide relevant and necessary data at the respective part of the industrial plant. The WPC communicates with the ADOMe of the respective object, to provide information for the next production step. Thus, produced objects can be registered early in the process flow.
Figure 2: Production scenario. different production lines, which are coupled via a workpiece carrier (WPC). The WPC accompanies the product through the milling, assembling and processing cycle and carries the product physically. The WPC is also equipped with a CPS- enabled ADOMe, that couples the physical product part with its virtual counterpart, which represents all product-specific data. Within this interconnected infrastructure, the WPC has access to all information of the product, to provide relevant and necessary data at the respective part of the industrial plant. The WPC communicates with the ADOMe of the respective object, to provide information for the next production step. Thus, produced objects can be registered early in the process flow.
Figure 3: Interaction of the individual components of the framework. The approach can be subdivided into three processing areas that need to interact with each other. Figure 3 shows the actual products and field device level, represented by each CPS and the associated ADOMe, furthermore, the supply level, where services, snippets and ADOMes are hosted as cloud- based networked solutions, and the assistance level for decision support and knowledge acquisition of the CPPS. Decision making is based on the dedicated processing steps and the context-adaptive provision of information of field devices and manufactured products stored in their ADOMes. Each product or field device has both a local and a global ADOMe. The local ADOM¢e is stored directly on the CPS with limited memory, and the global ADOMe, for storage-intensive data types, is stored by a central server, the OMS.
Figure 3: Interaction of the individual components of the framework. The approach can be subdivided into three processing areas that need to interact with each other. Figure 3 shows the actual products and field device level, represented by each CPS and the associated ADOMe, furthermore, the supply level, where services, snippets and ADOMes are hosted as cloud- based networked solutions, and the assistance level for decision support and knowledge acquisition of the CPPS. Decision making is based on the dedicated processing steps and the context-adaptive provision of information of field devices and manufactured products stored in their ADOMes. Each product or field device has both a local and a global ADOMe. The local ADOM¢e is stored directly on the CPS with limited memory, and the global ADOMe, for storage-intensive data types, is stored by a central server, the OMS.
Figure 4: Contextual management based on a situational model.
Figure 4: Contextual management based on a situational model.
Figure 5: OMS creates ADOMe in production. information sharing, and quality management. Within the cloud resources, such as processing power, memory or software, in the form of little scripts, are provided dynamically and appropriately over a network. The Object Memory Server (OMS) is the mair infrastructure component that stores ADOMes, according tc the model of an application server to serve a large number of users. Via a RESTful Web service interface the OMS permits access to process data of each manufacturer and provides the functions to create, store, replace, and modify the datz structures in a uniform and consistent manner. Figure 5 shows the interaction of the CPS’ client layer with the OMS. The
Figure 5: OMS creates ADOMe in production. information sharing, and quality management. Within the cloud resources, such as processing power, memory or software, in the form of little scripts, are provided dynamically and appropriately over a network. The Object Memory Server (OMS) is the mair infrastructure component that stores ADOMes, according tc the model of an application server to serve a large number of users. Via a RESTful Web service interface the OMS permits access to process data of each manufacturer and provides the functions to create, store, replace, and modify the datz structures in a uniform and consistent manner. Figure 5 shows the interaction of the CPS’ client layer with the OMS. The
Figure 6: Worker performing a maintenance task with a mobile device.
Figure 6: Worker performing a maintenance task with a mobile device.
TABLE I. Measurement statistic characteristics of a node at (1031.713. 2778.091)
TABLE I. Measurement statistic characteristics of a node at (1031.713. 2778.091)
Figure 5. The ROC curves of the OR fusion Figure 3. The histograms of sensor measurements
Figure 5. The ROC curves of the OR fusion Figure 3. The histograms of sensor measurements
Figure 7. The impact of number of cooperative sensors for OR fusion
Figure 7. The impact of number of cooperative sensors for OR fusion
Figure 6. The impact of number of cooperative sensors for AND fusion
Figure 6. The impact of number of cooperative sensors for AND fusion
To illustrate this principle, we take the example of the network represented in Figure 1. This network consists of 12 nodes numbered from 0 to 11. At the instant t, these nodes are arranged as described in Fig. 1. If we suppose that node 0 is the source and node 11 is the destination, the set of possible paths between the source and the destination at the instant t are described in Figure 1. Indeed, the possible paths are the following:
To illustrate this principle, we take the example of the network represented in Figure 1. This network consists of 12 nodes numbered from 0 to 11. At the instant t, these nodes are arranged as described in Fig. 1. If we suppose that node 0 is the source and node 11 is the destination, the set of possible paths between the source and the destination at the instant t are described in Figure 1. Indeed, the possible paths are the following:
Figure 2. Algorithm of paths construction in the case of periodic movements The instants of topology change:
Figure 2. Algorithm of paths construction in the case of periodic movements The instants of topology change:
P: Temporary variable containing the path which we are discovering. CP: Intermediate variable containing the current content o Figure 3. Algorithm of paths construction in the case of non-periodic movements
P: Temporary variable containing the path which we are discovering. CP: Intermediate variable containing the current content o Figure 3. Algorithm of paths construction in the case of non-periodic movements
-a- Variation of the maximum transmission delay in the simulations according to the moving radius of the nodes
-a- Variation of the maximum transmission delay in the simulations according to the moving radius of the nodes
-b- Comparison between the simulation transmission delays and the analytic transmission delays according to the moving radius of the nodes Bimire & Vanatnon af the waret craca tranemiceinn dalayv accaminna tn the analytic transmission delays according to the moving radius of the nodes Figure 5. Variation of the worst case transmission delay according to the moving radius of the nodes
-b- Comparison between the simulation transmission delays and the analytic transmission delays according to the moving radius of the nodes Bimire & Vanatnon af the waret craca tranemiceinn dalayv accaminna tn the analytic transmission delays according to the moving radius of the nodes Figure 5. Variation of the worst case transmission delay according to the moving radius of the nodes
Figure 4. Variation of the transmission delays over time
Figure 4. Variation of the transmission delays over time
BLE I. ASSIGNED FREQUENCIES FOR EACH HEXADECIMAL DIGIT Another researcher, Chung, proposed a near wireless control technology using high frequencies [12][13]. In this method, he defined high frequencies as base signals and low latency as a control signal. Base signals comprise more than two high frequencies and note the existence of the control signal to the smart device. Then, because low latency accesses various control data by frequency value, low latency is generated by the Central Processing Unit(CPU) of the smart device. The distance of the control signal in this method is greater than in Bihler’s method and the data transmission time is faster than in Lee’s method. Thus, Chung’s method can be used as a trigger signal for the effective transmission of advertising information.
BLE I. ASSIGNED FREQUENCIES FOR EACH HEXADECIMAL DIGIT Another researcher, Chung, proposed a near wireless control technology using high frequencies [12][13]. In this method, he defined high frequencies as base signals and low latency as a control signal. Base signals comprise more than two high frequencies and note the existence of the control signal to the smart device. Then, because low latency accesses various control data by frequency value, low latency is generated by the Central Processing Unit(CPU) of the smart device. The distance of the control signal in this method is greater than in Bihler’s method and the data transmission time is faster than in Lee’s method. Thus, Chung’s method can be used as a trigger signal for the effective transmission of advertising information.
Figure 1. Work flow of advertising method via smart device using high frequencies. First, the smart TV pictured on the left in Figure 1 starts sending the first high frequency signal to the smart device (@), and the smart device verifies the presence of the first high frequency using Fast Fourier Transform (FFT) (@). Then, if the smart device detects the first high frequency consistently, it checks whether the number of first high frequency signals is & times over (@)). Once this has been
Figure 1. Work flow of advertising method via smart device using high frequencies. First, the smart TV pictured on the left in Figure 1 starts sending the first high frequency signal to the smart device (@), and the smart device verifies the presence of the first high frequency using Fast Fourier Transform (FFT) (@). Then, if the smart device detects the first high frequency consistently, it checks whether the number of first high frequency signals is & times over (@)). Once this has been
Figure 3. Screenshot of the advertising application: (a) Main screen featuring a graph of the receiving high frequencies; (b) Real advertising information screen by advertisement signal. This section introduces the advertising application capable of supplying the advertising information to the smart device user by utilizing high frequencies as the proposed method. In addition, we explain the experiments and results used for the performance evaluation of the proposed method. We developed the advertising application based on iOS 6 and created it using Xcode 5. Figure 3 is a screenshot of the main advertising application.
Figure 3. Screenshot of the advertising application: (a) Main screen featuring a graph of the receiving high frequencies; (b) Real advertising information screen by advertisement signal. This section introduces the advertising application capable of supplying the advertising information to the smart device user by utilizing high frequencies as the proposed method. In addition, we explain the experiments and results used for the performance evaluation of the proposed method. We developed the advertising application based on iOS 6 and created it using Xcode 5. Figure 3 is a screenshot of the main advertising application.
Figure 2. Example of high frequency use in advertising information requests
Figure 2. Example of high frequency use in advertising information requests
Figure 4. Results of the test using each advertisement signal. Next, we evaluated the performance of the proposed method. We added an advertisement signal to the sound component of the TV content, which was a K-pop music video. We used 19kHz, 20kHz, and 21kHz as the first high frequency and 18kHz as the second high frequency. Each advertisement signal was generated twice at two-minute intervals. Thus, we did the receiving test 50 times for each advertisement signal. Because the advertising application checks high frequencies of FFT per 10ms, we set each threshold value in such a way that k was 12 times (60% of the first high frequency length) and i was six times (60% of the second frequency length). The distance between the smart TV and the smart device was 3m and the volume level of the smart TV was 70dB. The test space measured 5m x 4m, and 40dB was maintained as the silence space. Figure 4 below shows the result of the test using each advertisement signal. In Figure 4, when the first high frequency was 19kHz, the smart device received the advertisement signal 50 times. Using 20kHz as the first high frequency caused the signal to be received 49 times, 21kHz caused it to be received 47 times, and so on. Thus, the total success rate averaged 97%. This shows that transmission errors increase in line with the rising value of the first high frequency. We believe the reason for this is that the bin number of the higher frequency detected was lower than the bin number of the lower frequency over the same distance and with the same number of decibels. The next task was to test performance according to distance. The distance between the smart TV and the smart device was from 1m to 5m in Im intervals, and we did the receiving test 50 times with each advertisement signal (19kHz, 20kHz, and 21kHz) and at each distance. The test space and every threshold value were the same. Figure 5
Figure 4. Results of the test using each advertisement signal. Next, we evaluated the performance of the proposed method. We added an advertisement signal to the sound component of the TV content, which was a K-pop music video. We used 19kHz, 20kHz, and 21kHz as the first high frequency and 18kHz as the second high frequency. Each advertisement signal was generated twice at two-minute intervals. Thus, we did the receiving test 50 times for each advertisement signal. Because the advertising application checks high frequencies of FFT per 10ms, we set each threshold value in such a way that k was 12 times (60% of the first high frequency length) and i was six times (60% of the second frequency length). The distance between the smart TV and the smart device was 3m and the volume level of the smart TV was 70dB. The test space measured 5m x 4m, and 40dB was maintained as the silence space. Figure 4 below shows the result of the test using each advertisement signal. In Figure 4, when the first high frequency was 19kHz, the smart device received the advertisement signal 50 times. Using 20kHz as the first high frequency caused the signal to be received 49 times, 21kHz caused it to be received 47 times, and so on. Thus, the total success rate averaged 97%. This shows that transmission errors increase in line with the rising value of the first high frequency. We believe the reason for this is that the bin number of the higher frequency detected was lower than the bin number of the lower frequency over the same distance and with the same number of decibels. The next task was to test performance according to distance. The distance between the smart TV and the smart device was from 1m to 5m in Im intervals, and we did the receiving test 50 times with each advertisement signal (19kHz, 20kHz, and 21kHz) and at each distance. The test space and every threshold value were the same. Figure 5
Figure 5. Result of advertisement transmission test according to distance. below shows the results of the advertisement transmission test according to distance.
Figure 5. Result of advertisement transmission test according to distance. below shows the results of the advertisement transmission test according to distance.
Figure 2. Chording keyboard prototype used during the typing study
Figure 2. Chording keyboard prototype used during the typing study
Figure 3. The interface of the Java application used during the study A Java application was designed to display the text to be typed and to monitor the pressed keys. A screenshot of the application is shown in Figure 3. The top-left window contains the text to be typed and the bottom-left window represents the typing area. The help image is displayed on the right. Because we wanted to evaluate an error correction mecha- nism, we instructed the participants not to correct their mis- takes (however, this was not enforced and they could delete typed text). As a reward for the time commitment during the experiment, they received a fixed monetary compensation for the first nine typing sessions. To provide additional motivation, for the last session, the reward was proportional to the number of typed words and to the typing accuracy.
Figure 3. The interface of the Java application used during the study A Java application was designed to display the text to be typed and to monitor the pressed keys. A screenshot of the application is shown in Figure 3. The top-left window contains the text to be typed and the bottom-left window represents the typing area. The help image is displayed on the right. Because we wanted to evaluate an error correction mecha- nism, we instructed the participants not to correct their mis- takes (however, this was not enforced and they could delete typed text). As a reward for the time commitment during the experiment, they received a fixed monetary compensation for the first nine typing sessions. To provide additional motivation, for the last session, the reward was proportional to the number of typed words and to the typing accuracy.
Figure 1. Environmental progress and block.
Figure 1. Environmental progress and block.
The proposed work is justified by the importance of the teacher to follow, throughout the course, the implementation process, avoiding a posteriori analysis, i.e., with the results, the teacher is able to give special attention to students with difficulties in constant learning, directly, without requiring more than a tool for analysis. Introdugao a Integragao de Midias na Educagao
The proposed work is justified by the importance of the teacher to follow, throughout the course, the implementation process, avoiding a posteriori analysis, i.e., with the results, the teacher is able to give special attention to students with difficulties in constant learning, directly, without requiring more than a tool for analysis. Introdugao a Integragao de Midias na Educagao
Figure 1. Schematic representation of the 4 main artifacts of map matching applied to dense traces. For each case the artifact is represented to the left and contrasted with the ground truth trajectory, to the right.
Figure 1. Schematic representation of the 4 main artifacts of map matching applied to dense traces. For each case the artifact is represented to the left and contrasted with the ground truth trajectory, to the right.
Figure 3. Projected trace points plotted on top of the satellite map showing examples of Case-1: (a) before and (b) after match enhancement
Figure 3. Projected trace points plotted on top of the satellite map showing examples of Case-1: (a) before and (b) after match enhancement
Figure 5. Projected trace points plotted on top of the satellite map showing examples of Case-3: (a) before and (b) after match enhancement. Figure 4. Projected trace points plotted on top of the satellite map showing examples of Case-2: (a) before and (b) after match enhancemen
Figure 5. Projected trace points plotted on top of the satellite map showing examples of Case-3: (a) before and (b) after match enhancement. Figure 4. Projected trace points plotted on top of the satellite map showing examples of Case-2: (a) before and (b) after match enhancemen
Figure 6. Projected trace points plotted on top of the satellite map showing examples of Case-4: (a) before and (b) after match enhancement.
Figure 6. Projected trace points plotted on top of the satellite map showing examples of Case-4: (a) before and (b) after match enhancement.
Figure 7. Projected trace points plotted on top of the satellite map showing examples of Case-4: (a) before and (b) after match enhancement.
Figure 7. Projected trace points plotted on top of the satellite map showing examples of Case-4: (a) before and (b) after match enhancement.
Figure 1. INERTIA framework computational capacity) that are driven simultaneously to cooperate with each other as in Figure 1.
Figure 1. INERTIA framework computational capacity) that are driven simultaneously to cooperate with each other as in Figure 1.
Figure 4. Model view presenter
Figure 4. Model view presenter
Figure 5. Personal user interface d) HISTORICAL view: user monitors the historical data about consumption at different time; c) CALENDAR view: user submits to INERTIA world his habits and commitments such as for example “surgey in operating room from 11:00 am to 3:00 pm” or “training every Monday from 6:00 pm to 9:00 pm in the workout room”; All the components of the user interface architecture must communicate with each other. The connection between Presenter and View uses a socket connection. A socket is an end-to-end link over a single TCP connection, having the following features:
Figure 5. Personal user interface d) HISTORICAL view: user monitors the historical data about consumption at different time; c) CALENDAR view: user submits to INERTIA world his habits and commitments such as for example “surgey in operating room from 11:00 am to 3:00 pm” or “training every Monday from 6:00 pm to 9:00 pm in the workout room”; All the components of the user interface architecture must communicate with each other. The connection between Presenter and View uses a socket connection. A socket is an end-to-end link over a single TCP connection, having the following features:
580 California St., Suite 400
San Francisco, CA, 94104
© 2025 Academia. All rights reserved