Traffic Video Retrieval Based On Vehicle Features

Nowadays, with the fast development of economy and the expanding scale of city,the number of vehicles is much more than ever before, and many traffic problemsfrequently occur. This raises higher request for traffic regulators. Nowadays, publicofficial officers usually watch the traffic videos to find the video shots in which theillegal vehicles appeared. However, the long and tedious surveillance video does notreduce the traffic police’s working pressure and long-time watching cause visionfatigue which will result in false judgment and unpredictable loss. Facing with themass video information, how to analyze and retrieve the information we need isbecoming a hot issue.In this paper, we research on the vehicle detection process and introduce amethod on how to build an effective background model. Besides, we investigateseveral methods on remove the vehicle shadow. Then, we track the vehicles detectedbefore, and segment the occluded vehicles in order to track them separately. Weextract several vehicle features such as vehicle color, vehicle type and vehicletrajectory to represent each vehicle. We design an intelligent surveillance videoretrieval frame which supports query methods based on vehicle features. The maincontents of this paper are as follows:1. In moving vehicle detection, we research on some current moving objectdetection methods and propose a moving vehicle detection method based onbackground difference. The background is modeled based on samples. We model eachpixel with the information of former images. This algorithm has the advantages offaster speed, more robust to noise and stronger background adaptive capacity.As toshadow elimination, we suppress shadows in the detected moving regions withgradient filter according to the light and shadow analyze.2. In the aspect of vehicle tracking, we first introduce several target trackingmethods. Then, according to the real application scene, a method based on movingprediction is proposed. We can predict the approximate region of the vehicle in thecurrent frame according to the former vehicle region information; Then, match theregion with the connected regions that are detected in the last section, and the trackingprocess of this frame ends. Considering the occluded vehicles may have bad effect onvehicle tracking, a segmentation method is then introduced. This algorithm realizes multi-vehicle tracking. And our experiment shows the effectiveness of this method.3. In the extraction of vehicle features, we firstly assign each vehicle with an IDnumber. Then we can get the dominant color of the vehicle by color histogram analyze.The vehicle geometric character is classified into four types, and we get vehicle type.Finally, we get the information of the vehicle with trajectory analysis and clustering.4. In the aspect of frame design, this paper develops a traffic video retrievalframework based on vehicle features. In this paper, the features used for retrievalconsists of vehicle color, vehicle type and vehicle trojectory. We provide severalretrieval methods, including clolor based video clip retrieval, vehicle type basedretrieval and query by user’s sketch. It can return the retrieval results according tolow level features like vehicle color, vehicle type. It can also query by trajectorybased on user interaction. We can get the video clips that satisfy the user’s demand.The algorithms in this paper are tested in the experiment, and the experimentalresults show the algorithms used in the framework are effective.