Academia.eduAcademia.edu

Figure 1 – uploaded by Patrick Tchankue

See full PDFdownloadDownload figure
Several studies have found a strong correlation between bad weather conditions and driving errors [28]. The following reasons are often mentioned as a cause of these driving errors: a decrease in visibility, increased difficulty in controlling the car and decrease in the braking distance. When it is raining, roads are wet and become slippery. This results in an increase in the risk of car accidents, injuries and damage to the car. Figure 1. The definition of the driving context for this research

Figure 1 Several studies have found a strong correlation between bad weather conditions and driving errors [28]. The following reasons are often mentioned as a cause of these driving errors: a decrease in visibility, increased difficulty in controlling the car and decrease in the braking distance. When it is raining, roads are wet and become slippery. This results in an increase in the risk of car accidents, injuries and damage to the car. Figure 1. The definition of the driving context for this research

Related Figures (12)

Figure 2. The Architecture of Multimodal Interface for M obile Info-communication with C ontext (MIMIC) Framework
Figure 2. The Architecture of Multimodal Interface for M obile Info-communication with C ontext (MIMIC) Framework
Table 1. Classification of Driving Events The 10-fold cross-validation was used during the test phase; data was split into 10 equal sub-sets. The data was trained using
Table 1. Classification of Driving Events The 10-fold cross-validation was used during the test phase; data was split into 10 equal sub-sets. The data was trained using
Table 2. Confusion matrix for the IB1 algorithm (% ).
Table 2. Confusion matrix for the IB1 algorithm (% ).
Table 3. Confusion matrix for the J 48 algorithm (% ).
Table 3. Confusion matrix for the J 48 algorithm (% ).
Cae Seat a Tee a Participants drove on an urban road with | two lanes. The distance covered was about 6.5 kilometres with two right and left turns of about 90 degrees. The route had ten (10) speed bumps. Changes of lane were mostly recorded prior to take a turn.
Cae Seat a Tee a Participants drove on an urban road with | two lanes. The distance covered was about 6.5 kilometres with two right and left turns of about 90 degrees. The route had ten (10) speed bumps. Changes of lane were mostly recorded prior to take a turn.
Table 4. Confusion matrix for the IB3 algorithm (% ).
Table 4. Confusion matrix for the IB3 algorithm (% ).
Figure 4. Screenshot of the Capture of the Distraction L evel The data collection took place in two phases; the first phase was similar to the previous experiment. The driver and the primary researcher were the only occupants of the car. The primary researcher changed the distraction level by using the slider. The listraction level depends on the perception of the person capturing the data. Situations that require the driver’s full attention are associated with a high distraction level. The second scenario involved a vehicle with four passengers cluding the driver.
Figure 4. Screenshot of the Capture of the Distraction L evel The data collection took place in two phases; the first phase was similar to the previous experiment. The driver and the primary researcher were the only occupants of the car. The primary researcher changed the distraction level by using the slider. The listraction level depends on the perception of the person capturing the data. Situations that require the driver’s full attention are associated with a high distraction level. The second scenario involved a vehicle with four passengers cluding the driver.
Table 5. Performance of classifiers using the reduced training set
Table 5. Performance of classifiers using the reduced training set
Since the confusion matrices (Table 7, Table 8 and Table 9 showed good results, the prediction of the distraction level of the driver can be achieved using a model as described in Section 4.2. However the prediction of the driving context wil need some adjustment of the model. A designer can exclude some poorly recognised driving events such as “Speed Bump”, “Right Turn” and “Left Turn”. This driving context can be used to design a context-aware ICCS as illustrated by the Contex Adapter in Figure 2. When the driver tries to make or receive a call, the distraction level is calculated by the model. If the distraction level is Medium, Low or Very Low, the driving even model is queried. If the driver is not in a position to receive any incoming communication, a suitable adaptation mechanism should be implemented. These adaptations include delaying the incoming call notification, notifying the caller by sending a text message, sharing the driving status of the driver, or pausing the dialogue in case the driver was engaged in a conversation with the system.
Since the confusion matrices (Table 7, Table 8 and Table 9 showed good results, the prediction of the distraction level of the driver can be achieved using a model as described in Section 4.2. However the prediction of the driving context wil need some adjustment of the model. A designer can exclude some poorly recognised driving events such as “Speed Bump”, “Right Turn” and “Left Turn”. This driving context can be used to design a context-aware ICCS as illustrated by the Contex Adapter in Figure 2. When the driver tries to make or receive a call, the distraction level is calculated by the model. If the distraction level is Medium, Low or Very Low, the driving even model is queried. If the driver is not in a position to receive any incoming communication, a suitable adaptation mechanism should be implemented. These adaptations include delaying the incoming call notification, notifying the caller by sending a text message, sharing the driving status of the driver, or pausing the dialogue in case the driver was engaged in a conversation with the system.
Table 7. Confusion matrix for the KStar algorithm Table 7 show more details about the performance of the KStar algorithm. The highest confusion is the one between “Very High” and “High”, where only 92% of “Very High” distraction levels are correctly predicted. 7% were classified as “High”. Other classes had less errors; 97% of “Low” distraction was correctly predicted.
Table 7. Confusion matrix for the KStar algorithm Table 7 show more details about the performance of the KStar algorithm. The highest confusion is the one between “Very High” and “High”, where only 92% of “Very High” distraction levels are correctly predicted. 7% were classified as “High”. Other classes had less errors; 97% of “Low” distraction was correctly predicted.
The RandomForest algorithm seemed to work well with the lowest distraction levels. In Table 9, up to 98% of “Low” was As shown in Table 8, the IB1 algorithm also gave interesting results. The classes “Low”, “Very Low”, “Medium” and “High” were correctly classified by at least 95%. 91% of the class “Very High” was correctly classified while 7% were misclassified as “High”.
The RandomForest algorithm seemed to work well with the lowest distraction levels. In Table 9, up to 98% of “Low” was As shown in Table 8, the IB1 algorithm also gave interesting results. The classes “Low”, “Very Low”, “Medium” and “High” were correctly classified by at least 95%. 91% of the class “Very High” was correctly classified while 7% were misclassified as “High”.
Table 6. Classification of distraction level (94.49%). Conversely, the NaiveBayes and the SMO had the worst performance, respectively 51.86% and 65.84%.
Table 6. Classification of distraction level (94.49%). Conversely, the NaiveBayes and the SMO had the worst performance, respectively 51.86% and 65.84%.
Related topics:
Machine LearningDriver distractionIn-car communication system
580 California St., Suite 400
San Francisco, CA, 94104
© 2025 Academia. All rights reserved