Fig. 1. A layered approach to the semantic web. Fig. 2. Architecture of the ontology-based intelligent traffic light control system. The transportation domain is composed of several objects such as roads, junctions, vehicles, signals, etc. In this domain, there are many applications that should interoperate with each other. For doing so, a general specification reference including traffic light knowledge is needed to share and reuse it among the applica- tions. As a result of this necessity for Intelligent Transportation Systems (ITS) applications, an international standard called Geographic Data Files (GDF) specifies the conceptual and logical data model and the exchange format for geographic data bases (Lorenz et al., 2005). The ISO/DIS 14825(E) standard was repre- sented in 2002 with a concentration on ITS applications and emphasis on road and road related information. This standard encompasses the ontology of notions in the domain of traffic networks. As a result of extraction, extension and representation of this ontology in the form of OWL, the Ontology of Traffic Networks (OTN) has been created and used as a general-purpose ontology in ITS thoroughly (Oliveira and Camponogara, 2010). Nevertheless, it does not encompass traffic light control domain as good as other parts of traffic networks. Hence, OTN is not completely suitable for the purpose of traffic light controlling and we need an extended version of it. In the construction of TLCO, we have used two sub domains of OTN called Road and Ferries and Road Furniture and extracted some concepts such as Road, Road Element, Intersection, Traffic Light, etc. The rest of the TLCO’s elements have been defined based on literature and knowledge of experts such as traffic engineers and managers. In the first step of construction based on Methontology, the glossary of basic terms The architecture of ontology-based intelligent traffic light con- trol system has been adopted from layered structure of semantic web. This architecture contains three layers. The bottom layer is structured around imagery information extracted from images of intersection in the XML format. The middle layer is the traffic light control fuzzy ontology which represents the traffic light control knowledge including concepts, relations, properties, axioms and the rules in the form of OWL. At the top of the structure there is the reasoning layer which uses fuzzy ontology and XML instances for specifying real-time suitable sequence phasing and green times for each phase. Fig. 2 shows the architecture of an ontology based intelligent traffic light control system. The fuzzy traffic light control ontology as the most important part of the system is constru- cted manually and represents the traffic light control domain Traffic signal phasing fuzzy rules. Finally, the axioms and rules are defined based on related literature and standards such as HCM 2000 and GDF. We have defined some axioms and two types of fuzzy rules. For example, one of the axioms states that all left-turn movements are free. Hence, there are four main traffic movements in three ways and eight other ones in four-ways. Moreover, two types of fuzzy rules are defined for specifying phase types and optimized cycle time. These IF-THEN rules are based on some fuzzy input variables. The fuzzy variables for selecting suitable phase type are intersection type, the average vehicle and pedestrian congestion. The output variable, phasing type, is selected from a predefined set including Also for properties, some information including the amount of related concepts, value type, measurement unit, range of values and cardinality are specified. For instance, the “cycle time” Table 1 Fig. 4. Patterns for phasing types: (a) two simple phases, (b) lead-lag, (c) three simple phases, (d) four simple phases, (e) two composed phases, and (f) three simple phases and one pedestrian phase. For instance a fuzzy rule for phase type selection is stated as follows: “If an intersection type is four-ways, the average vehicle congestion is low, the average pedestrian congestion is medium, Then the phase type is simple two phases”. Also, the fuzzy rule Another fuzzy rule category is related to estimating the optimized cycle time. The fuzzy variables of these rules are weather condition, time, day and the average vehicle congestion. For this reason 60 fuzzy rules are defined. Weather condition is retrieved from the Meteorology Institute. Day and time are also Fig. 5. Three fuzzy sets for (a) day, (b) time and (c) cycle time. Images of an intersection contain semantic information in traffic light control domain such as road, road element, pedestrian crossing and divider position and also some properties such as vehicle and pedestrian congestion. This information is observable and understandable for humans but not for computers. Therefore, the images have to be processed and some machine understand- able description has to be extracted from them. For this purpose, we have used MATLAB image processing toolbox for region choosing to specify road element and pedestrian crossing location and for feature extraction including vehicle and pedestrian At the end of this step, the validity of TLCO’s items should be assessed. This function was done using expert’s opinions. The main purpose of the evaluation process is to show the usefulness of the developed ontology and its associated software environ- ment (GOmez-Pérez et al., 2004). Although all information, especially traffic light control rules has been extracted from international standards and scientific traffic literature, the knowl- edge of experts is the final validation criterion for correctness of them. In this process, all fuzzy rules including the optimized cycle time and phasing type rules, were prepared in the form of questionnaires and presented to some experts including three traffic colonels from Tehran Police Department and two traffic engineers from Tehran Traffic Organization. As a result of Delphi process of the expert’s opinion, about 84% of phasing type rules and 87% of optimized cycle time were accepted. Moreover, all TLCO’s segments including concepts, relations, properties and axioms were assessed and verified by these experts, simulta- neously. We evaluated the TLCO based on evaluation process of On-To-Knowledge methodology (Gomez-Pérez et al., 2004). Fig. 7. Reign choosing process. (a) An image of one way of an intersection; (b) the result of Sobel method running on the image; and (c) location of pedestrian crossing a1 divider on the image. Although these distances do not only contain road element boundaries and encompass some other sections such as pave- ments and trees, they are not a serious problem for congestion estimation and also the effects of this problem are decreased by applying neural network approach in the classification phase which is described below. As mentioned in the ontology construction section, pedestrian and vehicle congestions are fuzzy properties which take fuzzy values including low, medium and heavy. Because of quantitative and uncertain obtained values as a result of earlier activities, Fig. 8. Vehicle congestions in left and right road elements. Fig. 11. XML Schema for traffic light control instances. results from both of the trained networks, each of the output categories is classified in a different range. So, light, normal and heavy categories are classified as fuzzy sets shown in Fig. 10. Fig. 10. Fuzzy description for pedestrian and vehicle congestions. we use Kadlal basis runcuion neural NetWork (KDFNN) tO Classify the primary results in the categories that match the fuzzy membership functions. This network is a two layered feed forward network with useful properties such as simplicity, robustness and optimal approximation. It has successfully been applied to the image classification and image index (Chang and Fu, 2006).As mentioned before, obtained values for pedestrian and vehicle congestions encompass some noisy data that origi- nate from the nature of intersection environment such as the differences between lengths and widths of pedestrian crossing and road elements, existence of some unrelated objects in the road element areas such as leaves of trees and borders of streets which are different in the road elements. These differences cause uncertainty in the estimation of congestions. We solve this problem using another parameter which represents road ele- ment’s number. In this way, all road elements in the intersection are numbered from north and clockwise. This number along with related congestion is given to RBF neural network. This network consists of two layers including hidden radial basis layer and an output linear layer. The transfer function for a radial basis neuron is a Gaussian function ¢(x)=e-* and it is also a linear function for linear neuron. There are two different designs for RBFNN in MATLAB, one of which is referred to as the exact design (NEW- RBE) and the other one as the more efficient design (NEWRB). In this paper, the best results were obtained using NEWRB. Two RBF neural networks are considered for vehicle and pedestrian con- gestions in the form shown in Fig. 9. Both of them classify vehicle and pedestrian congestion in the light, medium and heavy classes which match values 1, 2 and 3, respectively. The image data set, consisting of 730 patterns, was classified by experts considering both vehicle and pedestrian congestions. =xs'schema xmins:xs="httpvwww.w3.org/2001 KMLSchema"> =xs:documentation xml:lang="en">intersection image schema</xs:documentation= =/xs:annotation= Traffic conditions of intersection. Table 2 Traffic parameters. Table 3 Fig. 12. Comparison between other approaches. Comparison between proposed approach and other approaches. Table 4 Table A1 (continued )
