Research And Application Of Multi-object Detection And Tracking Algorithm Based On Deep Learning In Video Surveillance

Multi-object detection and tracking algorithm based on deep learning is a technology that uses convolutional network to realize target detection and tracking.In the field of video surveillance,this algorithm can help the intelligent monitoring system identify and track the target more accurately,and improve the monitoring efficiency and accuracy.At the same time,it can be widely used in traffic safety,urban management,public safety and other fields.This paper mainly takes pedestrian multiobject detection and tracking in indoor scenes as the background,aiming at the problems of the current multi-object detection and tracking algorithm,such as missing detection and ID switching due to target occlusion and dislocation,as well as the slow speed and insufficient precision of target detection and tracking.By improving the network structure of YOLOX’s target detection algorithm and integrating the attention mechanism,the improved target detection algorithm is called YOLOX-P target detection algorithm.Based on this algorithm and ByteTrack target tracker,a pedestrian multi-target detection and tracking algorithm for indoor pedestrian target detection scenes is designed.The main work content of this paper are as follows:(1)Analyzing,reviewing and classifying comprehensively the object detection and tracking algorithms based on deep learning.Introducing and explaining in detail the famous one-stage object detection algorithm,YOLO series.Then analyzing the common problems of pedestrian object detection and tracking in video surveillance scenarios.Including Due to the large size transformation of the target,the small distance monitoring target and the occlusion and dislocation of the target caused by the problem of target omission,false detection and so on.(2)Aiming at the problems mentioned above,based on the latest YOLOX series algorithms,a series of improvement measures are proposed from the aspects of network structure and loss function,including introducing attention mechanism,adding CBAM to fuse attention module after backbone network output,replacing FPN as BiFPN,increasing feature extraction ability of input image,improving loss function,etc.DIoU Loss and Focal Loss were used to improve network accuracy and the improved network was named YOLOX-P.On the basis of the crowd monitoring data set,more than 1000 real monitoring images were screened to construct the data set and make the labels.Moreover,the ablation experiment of the improved YOLOX-P network was conducted on the data set and compared with the original network model to verify the effectiveness of the improved network.(3)Combing the improved YOLOX-P object detection algorithm with ByteTrack and training on MOT17 data set.Through comparison experiments with other detectors,the effectiveness of the improved YOLOX-P combined ByteTrack algorithm in the field of video surveillance was verified.