Cross-view Pedestrian Tracking In Wide-area Surveillance Video

Video surveillance constitutes an important part in public security,transportation and other fields.With the development of computer vision technology,video surveillance has attracted more and more attention from all sectors of society.As an important technology in video surveillance,it is of high theoretical and practical value to do research in cross-view pedestrian tracking.Cross-view pedestrian tracking in wide-area surveillance video was researched and divided into three sub-problems:pedestrian detection,single-view pedestrian tracking,cross-view pedestrian reidentification,whose algorithm implementations were studied respectively.In the section of pedestrian detection,a method based on deep convolutional network was studied.On the basis of a general object detection algorithm,the method of applying the algorithm to pedestrian tracking was proposed.In order to solve difficulties in pedestrian detection,the effects of three factors on the performance of detection algorithm were studied.Transfer learning was used to handle with the poor performance of cross-domain detection.In addition,image scale and priori anchor box were also adjusted respectively in order to improve the poor performance of small target detection.When dealing with pedestrian tracking,multiple target tracking and single target tracking were studied separately.In terms of multiple target tracking,the algorithm pipeline was divided into four steps: pedestrian detection,feature extraction,similarity computation and data association.In the step of pedestrian detection,the aforementioned detection algorithm was used.In the step of feature extraction,a handcrafted neural network and Kalman filter were used respectively to extract appearance feature and motion feature.In the step of similarity computation,Euclidean distance and IOU distance were utilized to compute similarity score between targets.In the step of data association,Hungarian algorithm was used to associate targets of same identity.In addition,a kind of Siamese network was proposed as the appearance similarity criterion to judge whether the target is same people before and after occlusion.When it comes to single target tracking,traditional single target tracking algorithms only use the first frame as the prior information for tracking,which fail to deal with various situations such as appearance change and occlusion.In response to this problem,the idea of tracking by detection was introduced,and Siamese Region Proposal Network was utilized as the base tracking model with IOU as the criterion to associate the detection target with the tracking target,which effectively improved the tracking performance.In terms of pedestrian re-identification,traditional algorithms assume that the query target must exist in the gallery database,which seldomly agree with the actual situation.Thus,a verification network which can check the similarity between the query target and the gallery target is proposed,thereby eliminating the dependence on the above assumptions.In addition,in order to improve the verification efficiency,a sampling method using temporal and spatial probability distribution was used to reduce the amount of calculation.Two problems were solved in this thesis: single-view multi-target tracking and cross-view single-target tracking.Experiments on multiple tracking datasets showed that the tracking methods proposed can effectively complete the two tasks.