Research On Object Tracking And Association Across Multiple Cameras For Intelligent Video Surveillance

With the development of digital image processing, pattern recognition,automatic control and computer vision, various video surveillance systems arewidely used in military and civil security. Multi-camera surveillance system hasthe advantages of wide range and big angle, and the association of movingtargets in multiple cameras is the basis for object tracking and matching inmulti-camera video surveillance. It is also one of the key techniques to improvethe intelligence of video surveillance system. Therefore, motion object trackingand association are the frontier research fields of intelligent video surveillance.At present, there are some effective object matching algorithms formulti-camera surveillance system. Among them, compared with thosepoint-based matching, the region-based matching does not need cameracalibration. However, it still has many problems in worthy of furtherinvestigation, such as the complexity of environment, object occlusions, thein-sufficiency of single feature in cope with the changes of environments, andhow to fuse multiple features for object matching. In this thesis, we aim at improving the accuracy of object matching and solving the difficulties caused byocclusion and illumination. The object matching is investigated for multi-camerasystem in order to improve the intelligent processing ability of video surveillancesystem, and the main contributions are summarized as follows:First, a feature fusion algorithm of region-based SIFT, color and geometricfeature is proposed for object matching, in order to solve the low accuracy inregion-based object matching by single SIFT feature, especially with theincrease of surveillance angle. The object detection is obtained by improvedbackground subtraction technique, and a particle filter is used for region-basedobject tracking. The SIFT descriptor is calculated for every object region, andthen fused with the color features and the geometric features of object contour.The object matching is obtained by comparing the distance of feature matrix.Experimental results show that the proposed method is robust under the sametype and different type object matching.Second, for multi-camera object tracking and matching withnon-overlapping fields of view (FOV), appearance matching is a valuable aidsince we have time and space constraints in systems capturing disjoint areas. A target matching method for multiple cameras by combining HOG and block LBPis proposed, and accuracy rate is computed by SVM. Using independent tracksof30different persons which were divided into three groups on average, wecompute the accuracy rate by experiments of HOG and LBP feature alone.Experimental results show that the proposed representation presents highresilience to incorrect object segmentation and illumination and it effectivelydiscriminates visual object. Average accuracy rate is up to80%.