Research On Multi-target Tracking Algorithm Based On Deep Learning

Multi target tracking is the tracking of multiple targets in a continuous video frame,that is,associating the same target in the front and back frames of the video and assigning a unique identifier.With the development and updating of deep learning methods,this technology is widely applied in fields such as intelligent security,traffic monitoring,and autonomous driving.However,multi target tracking tasks still face issues such as occlusion,motion blur,and difficulty in extracting target features,which will greatly affect the accuracy and accuracy of multi target tracking.In order to solve the above problems,this thesis proposes a tri-differentiation tracking method,and studies the feature extraction technology,target state prediction and feature matching methods of multi-target tracking.The specific research contents are as follows:(1)Aiming at the problems of false detection,ID switching and low training efficiency caused by insufficient feature performance,an improved channel attention multi-target tracking algorithm based on Fair MOT is proposed.The backbone network of this thesis uses the DLA(Deep Layer Aggregation)Network to extract the target features,and better integrate the semantic features and spatial features through the optimized jump connection.In order to enhance the representation ability of features,this thesis adds two channel attention modules: SE(Squeeze-and-Excitation)module and PSA(Pyramid Split Attention)module.By adding these two modules,the goal of improving model tracking performance and reducing training time is achieved,and by adding these two modules,the correlation between channels in the learning network is enhanced with only a small amount of computation.(2)Aiming at the problems of occlusion and motion blur in multi target tracking,a multi target tracking algorithm based on triple differentiation matching is proposed.Firstly,target scores under different occlusion and motion blur situations are obtained through the Center Net target detection algorithm.Then,the targets are divided into three categories: non occluded targets,slightly occluded targets,and heavily occluded or motion blurred targets.For these three types of targets,a combination of Kalman filtering,IOU(Intersection Over Union)distance measurement,and Hungarian algorithm is used to perform different schemes of target prediction and trajectory association.In this thesis,we add the correlation of the apparent characteristics of the high score target and the medium score target in the trajectory correlation part,and associate the high score target,the medium score target and the low score target with the trajectory in turn,which effectively reduces the interference between the targets,enhances the continuity of tracking,and alleviates the problem of identity switching of the target when occlusion occurs,so that the model can track all the detected targets more effectively.