Anti-occlusion Vehicle Tracking Algorithm Based On Kalman Filter

Intelligent transport system is the future of the transport system which can makethe traffic management more efficient. With the development of computer vision andimage processing, vehicle detection and identification based on vision is becomingmore potential. Vehicle tracking is a hot and difficult issue of intelligent transportsystem, which can provide vehicles’ trajectory, speed, traffic flow and other importantinformation, and lay the foundation for further analysis of vehicles’ behavior.Occlusion is an unavoidable problem in traffic video, which will have serious impacton vehicle tracking. Therefore, the main purpose of this paper is to study an anti-occlusion vehicle tracking algorithm.In order to do vehicle tracking, vehicles must be detected first. Vehicle detectionmeans to detect the changing area and divide the moving object from its background.The result of vehicle detection is directly related to the quality of vehicle tracking.Traditional Gaussian mixture model is improved by analyzing images’ entropy energy,and background can be updated at different rates based the images’ contents, so as todetect vehicles better. After vehicle detection, the foreground image must beprocessed, including image de-noising using label method and shadow detection usingcoarse model method. Vehicle regions will be clearer and conductive to vehicletracking after the foreground image getting rid of noise and shadows. Then connectedcomponent must be labeled. Pixels belong to the same connected region are given thesame mark, and different connected region are given different marks in order todistinguish each vehicle. Finally, vehicle region’s centroid position, area and othercharacteristic values must be extracted, so as to establish the vehicle’s parametermodel for vehicle tracking.A vehicle tracking algorithm based on Kalman filer is proposed in this paper.The Kalman filter is adopted to predict each region’s information in the next frame. Asearch region is created according to the predicted location. Searching match in thesearch region can avoid searching in the whole image, and help to improve thecomputing speed. The inter-frame vehicle regions’ relationship is established inaccordance with certain match rule. The relationship reasoning includes new object appearing, object disappearing, object sheltering and object separating. Theinformation predicted by Kalman filter is used to track vehicles when there wereocclusions. Meanwhile, the search region is enlarged to wait for the target appearingagain. If the vehicle has not yet appeared after a certain period of time, it will bediscarded. Experimental results indicate that the algorithm can get good trackingresults.