Research And Application Of Multi-Target Tracking Algorithm Based On YOLOV5 And DEEPSORT

The purpose of visual tracking is to estimate the action trajectory and activity state of the object in the subsequent video sequence,so as to provide data guarantee for the analysis and interpretation of the semantic content of the object.This paper is mainly based on the multi-target tracking field in the direction of target tracking in the field of computer vision.Simple Online Realtime Tracking with Deep Association Metric(DeepSort),a multi-target Tracking strategy based on detection,is studied.And applied to the tracking task.In this paper,combined with the target detection algorithm,a tracker based on YOLOv5 detection algorithm was selected to improve the DeepSort to improve the tracking effect of the algorithm on the pedestrian multi-target tracking data set MOT 16,and the further improved method was applied to the vehicle tracking data set UAVDT under UAV video.It has been proved that the improved algorithm also achieves good results in the field of vehicle video from the perspective of UAV.The research work of this paper mainly focuses on the following aspects:(1)Research on multi-target tracking algorithm based on YOLOv5 and DeepSORT.This part mainly studies the main strategy of multi-target tracking,compares the tracking effect of the traditional tracking algorithm and the multi-target tracking algorithm,and finally chooses the method frame based on detection,and improves the detection part based on DeepSORT.In this thesis,we try to use YOLO series detector to compare the performance of YOLOv3,YOLOv5,Residual Network(RESNET),Faster-RCNN detector and DeepSORT,and finally achieved better performance on the YOLOv5-S based DeepSort multi-target tracking algorithm.We evaluate the algorithm in MOT 16 dataset of pedestrian multitarget tracking,and reduce the index of identity change while keeping tracking performance.(2)Vehicle multi-target tracking applications based on YOLOv5 and DeepSORT.This part mainly aims at the YOLOv5-based multi-target tracking method proposed in the first part and further explores the scene application.It not only achieves multi-target tracking on the pedestrian tracking set,but also achieves good tracking results on the vehicle data set.In order to solve the problem of small targets with high density,the candidate area screening strategy in the detection part is optimized,and good vehicle tracking effect is obtained.(3)Multi-target tracking algorithm based on YOLOv5 and DEEPSORT is optimized for the number of parameters in vehicle tracking.This part is mainly aimed at the problem that the main multi-target tracking algorithm DEEPSORT takes a long time to train.The calculation amount of multi-target tracking is larger and the complexity is higher than that of the traditional target tracking,so the algorithm has higher requirements.On the premise of maintaining the accuracy,the number of model parameters is further reduced by modifying the network architecture to save the model training speed.It is verified in vehicle tracking that the algorithm after optimizing the number of parameters can also complete the multi-target tracking task well.