Research On Multi-Object Tracking Method Under Multi-Camera Collaboration

As people’s demand for automation and intelligent life increases,video-based object tracking technology has received widespread attention and application in intelligent video surveillance.Multi-object tracking technology has also become a highly anticipated part.In real-world multi-object tracking technology,due to the complexity of most scenes,multiple targets may intersect with each other or be occluded by other targets,which makes it difficult for multi-object tracking algorithms to identify and track targets correctly.Moreover,due to the limited monitoring field of view of a single camera,it is difficult to track targets across cameras.In this thesis,based on the detection-based multi-object tracking paradigm,we further improved the detection and tracking parts of multi-object tracking and optimized it by using multi-camera collaboration.The main work is as follows:(1)In order to solve the problem of poor detection and tracking performance caused by the limited field of view under a single camera,we overlapped the targets in the overlapping area of multiple cameras,stitched together the images collected by multiple cameras in the detection stage,and used it as the premise for network input.In order to improve the performance of the detector,we improved the variant DN-DETR of DETR,and used data augmentation to enhance the input data of the network to process the spliced images of multiple cameras.Finally,we replaced the original Res Net-50 with the better-performing Swin Transformer and optimized the loss function using Smooth L1 combined with CIOU LOSS to regress the position information of the detection boxes.The experimental results show that m AP＿0.5 has been improved by 0.178,and m AP＿0.5:0.95 has been improved by 0.150,both of which have been improved.(2)In order to solve the problem of tracking failure caused by target intersection and occlusion in the multi-object tracking process,we improved the Deep SORT algorithm.We used Mobile Net V2 to construct a re-identification network to extract deep features.In addition,we improved the IOU matching in data association by adding motion features to help the tracker better distinguish adjacent targets and handle changes in target motion.At the same time,introducing trajectory similarity for motion modeling further improved the accuracy of matching.The experimental results on the MOT15 and MOT16 datasets show that the improved Deep SORT algorithm has improved in both MOTA and MOTP.