Multi-objective Tracking Algorithm Based On Improved YOLOv5 And DeepSort

Multi-target tracking is an indispensable and pivotal task in the field of computer vision,which finds extensive applications in unmanned vehicles,video surveillance systems,and virtual reality environments.Due to the rapid development of deep learning,detection-based multi-target tracking algorithms can bring significant performance improvements to multi-target tracking tasks.However,multi-target tracking still suffers from object deformation,target occlusion prone to ID switching,etc.In this paper,YOLOv5 and DeepSort are used as benchmark models for improvement.The research work in this paper focuses on the following aspects:(1)The YOLOv5 model is used as the detection algorithm for the multi-target tracking method,and improvements are made for the YOLOv5 network.For the problem that small targets are easily missed during detection in the target detection task,it is proposed to add P2 small target detection layer;for the problem of object deformation and target occlusion during target detection,a deformable convolutional network is added to the backbone network;for the problem that target detection is easily interfered by complex background factors and so on,a SE attention mechanism is added to the backbone network,which can suppress the interference of irrelevant information and improve the Finally,starting from the loss function,Focal-EIOU is introduced to focus on high-quality anchor frames to obtain "high-quality" features and effectively improve the accuracy of regression.The optimized YOLOv5s model,named YOLOv5s＿PSF,was tested on the CrowdHuman dataset,and the accuracy of the improved algorithm was improved by 2.6%.(2)The improved YOLOv5s＿PSF model is used as the detection algorithm for the multi-target tracking method,and the DeepSort tracking algorithm is improved for the DeepSort tracking algorithm.Firstly,for the problem of poor prediction of irregular motion model by Kalman filter,the acceleration parameter is introduced to construct a uniform acceleration motion prediction model to improve the tracking accuracy of target motion state.Secondly,to address the problem that ID is easy to switch,EIOU is used as the distance association measure between the target detection frame and the trajectory tracking frame to solve the problem that the original IOU cannot reflect the distance relationship between the frames and enhance the trajectory matching process.Finally,testing on MOT16 dataset,the MOTA and MOTP of the improved multi-target tracking algorithm improved by 3.3%and 1.9%,respectively,and the number of ID switching decreased significantly,which in turn verified the effectiveness of the algorithm in this paper.