Research On Multi-Object Tracking Algorithm Based On Joint Detection And Tracking

With the rapid development of the computer industry and deep learning,object tracking technology has become more prominent in many fields such as traffic monitoring,assisted driving,smart medical care,and virtual reality.Among them,the pedestrian tracking algorithm that focuses on humans has become one of the hot issues in the field of object tracking.The object tracking algorithm estimate the appearance and motion state of the observed target by the information carried by video or image sequence.According to the number of tracking targets,object tracking can be divided into single object tracking(SOT)and multiple object tracking(MOT),and both of them can be subdivided into monocular and binocular object tracking according to the number of cameras.In this paper,the monocular MOT is studied.The current representative MOT algorithm of joint detection and embedding(JDE)and the MOT algorithm based on the fairness of detection and re-identification(Fair MOT)are improved.The specific research contents are arranged as follows :Firstly,this paper introduces the research significance of MOT algorithm and the various factors affecting the performance of MOT system.After that,the current research status at home and abroad in this field are introduced,as well as two main tracking ideas and related algorithms.Then,the traditional object detection methods are described,and leads to the object detection algorithm based on deep learning.The algorithm process and ideas of different types of detection models are expounded,and their pros and cons are analyzed in detail.At the same time,the theory and process of MOT are introduced,and the datasets and evaluation metric used in pedestrian MOT are explained.Secondly,based on the JDE algorithm,an online MOT algorithm is designed to balance the tracking accuracy and efficiency.The algorithm studies the influence of feature information loss on detection and tracking accuracy,and modifies the Darknet53 backbone network.Aiming at the problem of target information loss caused by too long path from low-level features to high-level features,a bottom-up path is added after the feature pyramid for feature fusion,and the obtained feature map is used for prediction.In order to improve the detection and tracking performance of small targets,spatial feature pooling is introduced to enlarge the receptive field and enhance the feature extraction ability.Subsequently,the transition layer is introduced into each residual module in the network to further optimize the repeated gradient information in the network and reduce the training cost.At the same time,considering that person re-identification only provides cropped pedestrian images,which is not conducive to MOT research.Therefore,the algorithm uses a method of reducing the embedding feature dimension to reduce the risk of overfitting caused by a small amount of datasets and improve the robustness of MOT models.Finally,the algorithm uses multi-task training method to train on a joint datasets of MOT17,CUHKSYSU and PRW.The test on the MOT17 test set reaches 71.9 % MOTA,and the speed fluctuates between 17.17 and 22.3 FPS according to the influence of input image size and pedestrian density.Finally,a pedestrian MOT algorithm based on high-resolution feature map is designed on the basis of Fair MOT algorithm.The main design idea of the algorithm is to improve the problems of pedestrians’ ID switching and target loss in dense pedestrian environment.The high-resolution feature map is used to predict the pedestrian keypoints,and the adaptive module is introduced to further improve the quality of keypoint regression.Then the algorithm uses a compromise heatmap estimation loss to balance the keypoints and non-keypoints.Ultimately,the ablation experiments of the optimized algorithm are carried out and trained under the joint datasets.Compared with the state-of-the-art algorithms on the benchmarks of MOT16,MOT17 and MOT20,it is proved that the optimization algorithm in this chapter can greatly reduce the number of pedestrian switching in dense scenes,and further improve the accuracy of the algorithm while ensuring real-time tracking efficiency.