Vehicles’ Image Features Extraction And Tracking Based On The Video Surveillance In The Intersection

Intelligent Traffic System(ITS) applies advanced technology(computer technology,communication technology, electronic technology, etc.) to traffic management, traffic controland so on to guarantee traffic’s safety, efficiency, environment and comfort. Car detection andtracking based on videos has been a research hotspot in ITS with the development ofcomputer vision technology. But most research is focused on multilane with the samedirection which has a simple scene. Research on car detection and tracking in the intersectionwhere the scene is complicated, traffic flow is heavy, occlusions and accidents happenfrequently is few. This paper focuses on research on car detection and tracking based on thesurveillance video in the intersection.Cars detection and tracking algorithms based on surveillance videos are studied byusing the intersection surveillance video and analyzing features of the intersection trafficscene.First, the common moving objects detection algorithms are introduced. The backgroundsubtraction is chosen by analyzing the pros and cons of detection algorithms. Then, threecommon background modeling algorithms are implemented and tested using CDnet database.The detection results of three algorithms under different challenging scenes(camera jitter,dynamic background,etc) are calculated. Results show that Gaussian Mixture Models(GMM)can maintain high detection rate and precision in varied challenging scenes which means thatGMM works for detecting outdoor moving objects. But results also prove that GMM hasfollowing shortcomings: objects in the shadow area can’t be detected entirely and objectswhich stay still for too long are detected as background. In view of the above shorts, twoimprovements are put forward: the first one is that a different matching coefficient is used inthe shadow area which is detected using threshold method to enhance detection rate inshadow area; the second one is that background model updating strategy which only updatesthe background pixels is chosen in case of missing still cars’ detection. Experiments provethat improvements have increase the detection rate. a improved tracking method based on theKalman Filtering is proposed according to features of intersection which include the heavytraffic flow and frequent occlusions. During tracking, two different matching techniques areused and corresponding relationships between tracks and detections are judged by using theintersection method. Experiments prove that the tracking method proposed in this paper canreduce the tracking mistakes caused by occlusions and running in and out of the camera.Research in this paper can be helpful in promoting the application of cars detection andtracking technology in ITS and promote the development of traffic data collection, trafficincident detection, etc.