Research On Multi-Pedestran Tracking Algorithm In Video Surveillance Scene

With the popularization of artificial intelligence technology,pedestrian multi-target tracking technology has been widely applied in video monitoring,medical assistance and automatic driving.However,in the process of pedestrian target tracking,the image information input to the tracker changes as the target moves,and the target is also affected by occlusion and background clutter during the tracking process.Therefore,how to more accurately determine the location and identity information of the tracked target is still a problem that needs to be studied in depth in the field of vision tracking.Starting from intelligent video surveillance,this paper studies and improves pedestrian tracking algorithm based on YOLO detection model and JDE tracking model.The main work is as follows:(1)A feature extraction network optimized by YOLO deep learning model was proposed to solve the problem that the pedestrian detection task and the pedestrian re-recognition task were competing and the algorithm could not estimate the pedestrian position and range accurately in the process of pedestrian multi-target tracking.Based on the concept of multi-task learning,a task decoupling structure is designed.The structure adjusts the detection feature by re-recognition feature.The structure can learn the common features of the two tasks,so as to reduce the competition relationship in the task learning process,and increase the performance of the detection task.Then,attention is paid to the feature channel through a sensory model that introduces attention mechanism.For the attention feature map of the attention part,we obtain it through Meanpool,Conv and Sigmoid.The experimental results show that the tracking accuracy and tracking effect of the tracker are both improved after applying the above optimization method.(2)To address the problem that it is difficult for the tracker to accurately match pedestrians between frames in the case of frequent pedestrian movement and pedestrian occlusion that may be encountered in the pedestrian multi-target tracking process.In this paper,a pedestrian appearance feature fusion structure is designed to improve the discernibility of pedestrian appearance features by fusing three different depth features.At the same time,the feature fusion method is optimized in the pedestrian feature matching stage,and the prediction results of pedestrian target motion feature are improved by kernel correlation filtering algorithm.The improved fusion feature can effectively reduce ID variation and improve tracking accuracy in the path connection of pedestrians with multiple targets.(3)A pedestrian multi-target tracking system for surveillance video is completed.To address the problem of possible degradation of image information quality in practical applications of video surveillance,the tracking system preprocesses the surveillance images by limiting the contrast adaptive histogram equalization method.From the experiments,it can be seen that the tracking accuracy of the tracker is improved after the image pre-processing operation,and the visual effect of pedestrian multi-target tracking is further improved.