Research On Pedestrian Tracking In Intelligent Video Surveillance

The purpose of video surveillance is to endow the computer vision system with human’s recognition ability to detect, track and comprehend the state of the object. As the core technology of security field, the video surveillance is getting popular in recent years. The pedestrian tracking is a base and core of video surveillance developed over decades, but detector has some blemish such as low precision, object lost and identify switch et al which due to the effect of the variety of pedestrian motion and illumination as well as the occlusion.The critical objective of the thesis is to research the robust pedestrian tracking algorithm in camera system. The contents of the thesis cover the tracking of single objcet single camera, multi-object single camera and multi-object multi-camera. Besides, the related methods are studied such as object detection and keypoint matching.The proposed algorithms are summarized as follows:(1) We present a high performance detector which included a discriminative LSSS descriptor based on the saliency. The descriptor represents the local shape feature of pedestrian in polar coordinates, and then Real AdaBoost algorithm is used to form a simple and effective strong classifier. This method is sample and improves the ability of dealing with the occlusion, which provides the perfect results for tracking-by-detection algorithm.(2) We study the single camera tracking algorithm. Firstly, to resolve the lost problem, a novel single object single camera tracking algorithm combining projection histograms with a centroid shift is proposed. The projection histograms are spatio-colorimetric presentation comparing with the classic color histograms, and the centroid shift is used instead of the resample process of the particle filter, it improves the ability of tracker to deal with part occlusion. Secondly, to decrease the phenomenon of label switch, the group based single camera multi-object tracking algorithm is proposed. The tracking problem is converted into minimum the energy function which combines the conditional random field mode and label cost. Experimental results demonstrate that the proposed algorithm decreases the probability of label switch and improves the robustness of the algorithm comparing with the classical methods.(3) The dissertation study the multi-camera tracking algorithm. Firstly, to obtain the projection relation between cameras, the approach presents a binary descriptor matching algorithm based on hierarchical learning method. The descriptor learning process is divided into two levels of coarse and fine, which combines the advantages of the fixed-point sampling mode and random sampling mode. It enhances the performance of learning. Secondly, the proposed work presents a multi-camera multi-object tracking algorithm based on weighted consensus, it combines the target’s color feature with the projection relation in network flow framework and enhances the power of dealing with occlusion. The weight consensus algorithm modifies the error caused by detector and camera calibration, therefore, it improves the tracking precision.