Research On Multi-object Tracking Algorithm Based On Deep Learning

Multi-Object tracking(MOT)is the task of stably tracking multiple objects of interest in an image sequence and accurately determining the position of each object at each time step.MOT has important applications in areas such as autonomous driving and security systems.Currently,deep learning-based multi-object tracking methods can be mainly divided into two categories based on whether detection and tracking are independent: Detection Based Tracking(DBT)and Joint Detection Tracking(JDT)methods.Among them,DBT requires object detection first,followed by tracking.JDT integrates the tracking part into object detection,which has advantages in time efficiency compared to DBT.The Joint Detection and Embedding(JDE)method in JDT simultaneously completes object detection and reidentification tasks,achieving a balance between speed and accuracy,and attracting more attention from researchers.However,JDE methods currently face some challenges,such as the balance problem between object detection and re-identification tasks,and the feature matching errors problem in data association.Therefore,this thesis focuses on the joint detection embedding model and data association strategy in the one-shot method.The main research contents are as follows:(1)A multi-object tracking algorithm based on Re ID feature enhancement is proposed to address the conflict between object detection and re-identification in the joint detection embedding model.The conflict arises due to the expectation that the detection network reduces intra-class variance to give the same semantics to objects of the same class,while the reidentification task is more concerned with the differences between different individuals within the same class of objects,hoping that the network can amplify these differences.Therefore,a task-correlated attention network is used to reconstruct the feature maps sent to each task branch using self-attention mechanism,so that each branch obtains task-related feature maps.Meanwhile,a smooth gradient-enhanced cross-entropy loss is proposed to gradually shift the focus to hard negative samples in the training process,enhancing the quality of extracted Re ID features.Experimental results on the MOT20 dataset show that the IDF1 score reaches 68.9%and the MOTA reaches 63.0%,which is a 1.6% and 1.2% improvement compared to popular single-shot methods,proving that the proposed method improves tracking accuracy by enhancing the quality of Re ID features.(2)A multi-object tracking algorithm that fuses feature measurement is proposed to address the ID switch problem caused by unmatched features in the current data association strategy due to target occlusion.Weighted fusion of motion features and appearance features are used to measure the similarity between the current target and the trajectory.The similarity of appearance features is measured using cosine distance,and the similarity of motion features is measured using Efficient Intersection over Union(EIo U).In addition,a second-order association method is used to match high-scoring and low-scoring detection boxes,reducing the loss of real targets,and using fusion feature measurements to filter out background detection boxes.Experimental results on the MOT20 dataset show that the IDF1 score improves by 1.6%,the MOTA improves by 5.2%,and the IDs decreases by 2781,indicating that the improved data association strategy can effectively solve the ID switch problem caused by occlusion.