Theory And Methodology Of Mixed Traffic Data Collection For Pedestrians And Cyclists Based On Image Processing

Among the mixed traffic system, the traffic flow produced by pedestrians and cyclists has severely influence on the vehicles. At present, this influence becomes a major threat to the traffic safety, and could as well lead to the increasing delay and decreasing capacity in the traffic network. In this case, the major issue for the improvement of current traffic situation lies in the active controlling of pedestrians and cyclists, as well as effective improving of the traffic capacity in the urban traffic network especially at the signalized intersection, leading thereby to less travel time and higher safety level for all the travelers. Thus, more and more Intelligent Traffic Systems (ITS) have been developed and implemented in the traffic management and control practice. As the first primary element in ITS, the traffic data collection devices are playing an increasingly important role in many types of ITSs.Considering the above perspectives, the collection of traffic flow data for mixed traffic objects would be meaningful for in-depth analysis of typical mixed traffic circumstance in China, thus providing theoretical basis for the construction, post-management, and evaluation of traffic infrastructures. This study mainly focuses on the data collection of pedestrians and cyclists with image processing techniques as the major analytical tools. The main content of this dissertation is summarized as follows:(1) Summarizes the development of current traffic data collection system, and the existing theories and methods in vision-based system. In addition, the limitations of current methodologies on collecting mixed traffic data has been analyzed in details. Based on the foregoing conclusion, this study establishes a theoretical framework for mixed traffic data collection including the following four modules: object detection, object tracking, feature extraction, object recognition.(2) Considering the difficulties in accessing the reliable background image in real traffic scenario, this study makes comparison on the three models of Running Average, Temple Median Filtering and Gaussian Mixture Model, and establishes a self-adaptive background extraction model based on mathematical morphology. This model could extract the background image at a higher precision level and preserve more primitive information in the original image. The experimental test indicates that this model has higher precision than the other three, and could operate fast enough to meet the requirement of realtime processing. Beside, this model could also self-adaptively adjust the brightness of image according to the changing of ambient lighting. This ensures the extracted background to remain at stable brightness, which is more close to the reality.(3) In the part of object detection, this study concerns various factors with improving the robustness of the algorithm, including: the selection of self-adaptive threshold; the noise suppression technique based on neighborhood information, light intensity and color information. Among which, the neighborhood information could effectively help eliminate the system noise, while the light intensity and color information could help remove the environmental noise.(4) In the part of object tracking, this study proposes a fuzzy-based Kalman Filter model dealing with the object overlapping, gathering and scattering etc. The model matches the prediction value produced by Kalman Filter to the observed value based on fuzzy theory. This method ensures that the moving objects would be tracked precisely under several circumstances of object occlusion and data missing. Thus, it is more suitable for the object tracking in mixed traffic scenario.(5) Based on the analysis of binary image for pedestrians and cyclists, this study combines the pixel distribution and morphologic characteristic of binary image as the input vector of BP neural network, leading thereby to great improvement in the recognition precision.