Researching Of Single/Multi-target Pedestrian Tracking Algorithm In Fixed Background

Intelligent video surveillance is the research focus in the field of computer vision in recent years, and has a broad application in the field of security, intelligent transportation and behavior analysis. Pedestrian target detection and tracking is the key technology of intelligent video surveillance, and it is also the research foundation of subsequent target identification and behavior analysis. This paper will research the single target tracking problem with complex background and multi-target tracking problem with fixed background.Target detection is the basis for tracking. This paper first analyzes the advantages and disadvantages of three commonly used target detection algorithms. Considering the specific application environment, background subtraction method with Gaussian mixture background model is selected to detect foreground objects. Finally, complete foreground objects will be extracted through shadow removal, morphological filtering and connected component analysis.For single target tracking problem, in order to improve the robustness of the tracking algorithm, we proposed an adaptive multi-feature integration single-target tracking algorithm based on particle filter algorithm. First, select the tracking target manually. By comparing the ability of distinguishing between the target and the background of a variety of color and texture features, we select the two optimal features to describe the target. Taking the two selected features as two different target models, two estimation results of target position will be obtained by using particle filter algorithm. If the two estimation results are similar, it means the selected features are active and the tracking result is reliable. If the two estimation results are quite different, it means one or two selected features are unreliable and the tracking result has a large offset. For this case, the reliability of estimated result of last frame will be used to decide whether to return the last frame to reselect optimal features and do particle filter estimation in current frame again. In order to ensure the accuracy of the target models. the target models will be updated only when the two estimated results are similar.For multi-target tracking, there may be a large number of targets in the scene and the states of the targets may be complex. Considering the real-time nature of the algorithm, the multi-target tracking algorithm will be based on region detection in this paper. First, the background subtraction method based on Gaussian mixture background model will be used to extract foreground objects. Then, according to the coincidence between the foreground objects in current frame and the tracking targets in last frame, a correlation matrix will be constructed between them. According to the relationship between the rows and columns, target state can be grouped into five categories:target appear, normal state, target fusion, target split and target disappeared. Different target state corresponds to different tracking algorithm. Because the reasons for target split are more complex, this paper analyzes the reasons which may cause target split and proposes the corresponding processing methods. Especially for the split after the integration of multi-target, we propose a tracking algorithm based on color matching.Experimental results show that the single target tracking algorithm is better in the case of the dramatic changes background and similar objects interference, and the multi-target tracking algorithm can effectively deal with the special case of the multi-target fusion and split.