Research And Implementation Of Multi-object Tracking System For Video Surveillance

Multi-object tracking for video surveillance is one of the key technics in computer vision and artificial intelligence which has gained increasing attention due to its tremendous scientific research and commercial value.Although numerous approaches have been proposed to solve this problem with demonstrated success,it remains a challenging problem in complex scenes,because of frequent occlusion,similar appearances,and other factors.From two aspects,i.e.,multi-object detection and multi-object tracking,this dissertation further analyzes the basic theory and key technics,and compares the advantage and disadvantage of conventional approaches and mainstream methods,respectively.Our approach adopts the association based tracking model to solve the multi-object tracking problem by associating tracklets in different ways according to their confidence values.In order to improve the association precision and discriminate the appearances of objects,an incremental linear discriminant analysis is used for updating appearance model.Based on the above tracking algorithm,this thesis proposes to segment the occluded objects into numerous superpixels,cluster the segmented superpixels for constructing appearance models and computing a target-background confidence map to distinguish the occlusion order of mutually overlapped objects,then re-establish the appearance model of occluded objects.For the purpose of solving the failure of tracking due to severe occlusion,this thesis puts forward a method to track the local appearance of object by seeking the superpixels of high confidence in areas where occluded objects may exsit.Extensive experimental results demonstrate that the proposed multi-object tracking algorithm can accurately track multiple targets and show better performance under severe occlusion in video surveillance.