Research On Pedestrian Tracking Algorithm In Video Surveillance

In the era of big data,more and more data needs to be processed in different industries,especially video data.Prior to the advent of video algorithms,video data was processed manually.Obviously,manual processing and analysis of video data requires a lot of manpower and is inefficient.The content of this paper is based on the multi-target tracking algorithm in pedestrian surveillance video.We optimize the detection module and pedestrian feature extraction module in the multi-target tracking algorithm framework,and select the optimal lost trajectory survival time.Finally,a visualized system for monitoring video detection and tracking is implemented.The main work is as follows:(1)In order to optimize the object detection module and obtain a detection algorithm with high speed and accuracy,this paper proposes a lightweight Hourglass backbone based on the Center Net detection algorithm,which improves the efficiency of the algorithm without losing accuracy.Specifically,the module of the Hourglass-104 convolution layer is optimized,the size of the convolution kernel is modified,the dimension of the feature map in the convolution layer is reduced,and the original heavy structure is optimized as a lightweight structure.The improved algorithm is tested on the MS COCO data set,and the comprehensive performance is ahead of the classic algorithm.(2)In order to optimize the feature extraction module in the tracking algorithm,after full investigation,we selected the PCB-RPP network that is more prominent in the field of pedestrian re-recognition for research.The feature extraction mechanism is studied,including feature hard partitioning,soft partitioning mechanism and the number of blocks of the extracted features,and the MOT16-Re ID700 pedestrian re-identification dataset is created based on the MOT16 dataset.In the MOT16-Re ID700 and Market-1501 data sets,the PCB-RPP network was tested for different numbers of blocks.Through experiments,it is determined that when the PCB-RPP network extracts features for pedestrians,the number of blocks is 6.(3)In view of the situation that the comprehensive performance index MOTA(Multiple Object Tracking Accuracy)of the Deep Sort multi-object tracking algorithm is not high enough and the ID switch is large,this paper proposes to use the pedestrian part-level feature extracted by PCB-RPP network to replace the original depth feature,and adopts the improved detection algorithm of this paper to provide the detection results for Deep Sort.The trajectory survival time is extended,and when the target appears again after being blocked for a short time,the ID switch situation is avoided.(4)Combining the multi-target tracking algorithm with computer front-end and backend technology,realizes the visual display of real-time detection and tracking of surveillance video,which can be viewed on the browser.