Research On Vision-based Road Traffic Parameter Extraction

The total number of vehicles is increasing in every city due to rapid economicdevelopment. As a result, it is a hard task to solve the problem traffic congestion brings.Currently, a variety of traditional traffic detectors have been installed in kinds of roadsby the traffic management department. However, the traditional traffic detectors can’t bewidely applied to traffic control management because of the their own shortcomings,such as high cost, monotonous traffic parameter gathered, low precision, little detectionrange. Nowadays, it has been an important development direction of IntelligentTransportation Systems (ITS) to use computer vision technology in processing andanalyzing traffic videos from traffic surveillance cameras installed in the roads, in orderto obtain parameters of traffic flow, traffic speed, vehicle types, traffic occupancy, etc.In this paper, some key technologies are studied in the video-based trafficparameter acquisition system, such as vehicle detection, vehicle tracking, vehicleclassification, road traffic state judgment.Firstly, in the part of vehicle detection, several common moving vehicle detectionmethods are introduced and analyzed. Finally, an improved GMM algorithm based onEM is adopted in detecting and segmenting the moving vehicles. Experimental resultshows that the algorithm can detect and segment vehicle objects well in a variety ofoutdoor conditions.Secondly, a novel method is proposed to improve real-time of system to detectvehicle objects in the complex environment. In my method, the computation is reducedand the interference of moving objects is excluded from the non-road areas by ROI areato set the detection section. In addition, the method of edge feature fusion is used inremoving vehicle shadows under the bright light, as a result, it is effective to avoid falsedetection for the non-shaded areas instead of the traditional method to detect shadowsonly based on shadow properties in the HSI color space. Then, the color imagehistogram equalization is adopted to improve image contrast and enhance the efficiencyfor detecting vehicles in foggy weather and dark conditions. Besides, a simplesegmentation about dynamic occlusion problem of vehicles also is introduced in thispart.Thirdly, in the traffic flow detection part, the tracking method based on Kalmanfilter is adopted in tracking the vehicles detected. Simultaneously, the result of vehicle tracking is also used in counting the number of vehicles detected.Fourthly, in the part of vehicle classification, the geometric features are extractedin the moving vehicle section in video. Then, the vehicle classifier is designed based onBP neural network. The parameters of the neural network are trained by the sample dataset given. Thus, the vehicles from the urban roads can be classed into three types by theneural network trained, that is, the large vehicles, small cars and motorcycles.Finally, in the part of judging road traffic state, the corresponding methods ofcongestion status evaluation are proposed according to two different cases, where one isthat the moving vehicles can be spitted out, and the others can’t.