Research On Multi-object Tracking Methods Based On Trajectory Information

Multi-Object Tracking(MOT)tasks aim to track multiple moving objects in video or image sequences,accurately identifying and recording their positions and motion states in each frame.It has a wide range of applications in fields such as security monitoring,virtual reality,and autonomous driving.With the rapid development of object detection technology,most existing MOT methods follow the tracking-by-detection strategy,which divides the MOT problem into two independent tasks:object detection and data association.These methods mostly perform object detection on individual images,considering only the visual clues from a single frame,and assume that all objects in a single frame can be accurately localized by the object detector.This overlooks the negative impact of various real-world challenges,such as severe occlusion and motion blur,that arise from the video time dimension.This indicates that relying too much on single-frame detection is unreliable.To address these problems,this paper explores the spatiotemporal information in the video stream from a spatiotemporal perspective,continuously perceiving the temporal coherence of moving objects while considering the current visual clues.Based on advanced object tracking technology,this paper proposes a MOT algorithm,which can be summarized into three parts:(1)To further improve the ability of multi-object tracking methods to deal with complex scenes,this paper proposes a trajectory information reuse strategy from a temporal perspective,which mines the effective information in historical trajectories and preserves it in the form of trajectory masks,and uses trajectory masks to assist in multi-object tracking in future frames.Combining the trajectory information reuse strategy with a convolutional neural network-based object detection network and data association method,this paper proposes a trajectory mask-based multi-object tracking algorithm(TMTrack).Firstly,the trajectory mask generation algorithm is designed to save the available spatiotemporal information in historical trajectories.Secondly,the trajectory mask network is designed to extract features from trajectory information and fuse them with the backbone network features.Then,target position prediction is carried out,and finally,data association is performed on the generated detection boxes.The algorithm effectively integrates object detection,data association,and time clues into a unified framework,achieving end-to-end optimization.(2)As TMTrack adopts a convolutional neural network as the feature extraction network for processing image data,the object detector based on the convolutional neural network cannot handle global information and long-term dependencies when processing image data.At the same time,when processing image data,the convolutional neural network requires a large number of convolutional and pooling layers.As the network becomes deeper and wider,the number of model parameters will become very large,which is prone to overfitting and long training time.Therefore,this paper introduces an advanced Transformer backbone architecture to rebuild the TMTrack tracker’s backbone network and redesigns a lightweight trajectory mask network to adapt to the new backbone network.The performance of the optimized object tracker has been improved in terms of inference speed and tracking accuracy.(3)Compared with other excellent multi-object tracking methods,TMTrack has shortcomings in maintaining target identity,which can cause more identity switches.In this regard,this paper analyzes the reasons for this situation and further optimizes the TMTrack method.First,the trajectory mask generation algorithm is optimized to aggregate the trajectory information of multiple frames.Second,appearance measurement is introduced in the data association stage to match trajectories and detection boxes by combining motion model measurement and appearance measurement,in order to improve the accuracy and robustness of multi-object tracking,making it more suitable for complex scenarios.Multiple groups of ablation and comparison experiments were conducted on benchmark datasets,verifying the effectiveness of the proposed multi-object tracking algorithm.