Research On Multi-target Tracking Algorithm Based On Deep Learning

In recent years,multi-target tracking has been a hot topic in the field of computer vision.The multi-target tracking network model can obtain the exact position of each target in the video frame and continuously track each target.However,in practical application scenarios,it is inevitable that the targets are too dense and crowded and the targets become smaller from near to far and the targets overlap and obscure each other,which will make the performance of multi-target tracking poor,the accuracy is not high,the target labels switch frequently,and then the targets cannot be tracked accurately without interruption.The main research objective of this paper is pedestrian multi-target tracking,to improve and optimize the existing detection algorithm and the shortcomings of the tracking algorithm,in order to improve the tracking accuracy of the whole model in the case of dense and crowded targets and smaller targets and to reduce the number of label switching when the targets reappear after being occluded,the main work and innovation of this paper are as follows:(1)In this paper,the Mask R-CNN target detection network model is improved.The backbone network of Mask R-CNN network model is replaced with ResNeSt101 network to improve the accuracy and precision of the model in the case of smaller target detection.(2)In this paper,an iterative target detection scheme is added to the Mask R-CNN network model.The model performs one more target detection after the first target location detection,so that the targets not detected in the first time can be detected and more targets can be detected.(3)In this paper,we propose an improved multi-target tracking algorithm based on SORT.Firstly,the target color histogram feature is added to the Hungarian algorithm,so that it not only considers the motion features of the target,but also uses the apparent features of the target to match two adjacent frames of targets when performing feature matching;then the data cascade matching algorithm is improved to save the disappearing targets as well as the newly appeared targets as indeterminate trajectories and give survival time to reduce the number of target label switching.By replacing the backbone of the Mask R-CNN network model with ResNeSt101 network and adding an iterative target detection scheme,the feature extraction capability of the network model and the detection accuracy of targets in dense scenes can be effectively improved;meanwhile,the improved SORT multi-target tracking algorithm can reduce the number of label switching of targets.The experimental results show that the improved Mask R-CNN network model studied in this paper improves the recall rate by 12.8% and the average detection accuracy by 12.2% compared with the original model,improving the detection accuracy of small targets and the recall rate of targets;meanwhile,the improved SORT-based multi-target tracking algorithm reduces the number of label switching by 875 times compared with the SORT multi-target tracking algorithm,and the tracking accuracy is improved by 3.6% and tracking accuracy is improved by 1.6%.These results show that the improved method proposed in this paper can effectively improve the tracking efficiency of the model in dense scenes,enhance the tracking accuracy,and provide advantages for the tracking of multiple targets in dense scenes.