Feature Enhancement And Data Association For Multi-Object Tracking

By processing the video sequence,Multi-Object Tracking(MOT)aims to acquire and evaluate the appearance features,bounding box,and motion states of all objects in each frame of images,and generate a continuous trajectory of each object.As a result,the online MOT technology has a lot of potential in the video sequence.Due to the complexity of surveillance scenes and the interactive occlusion of dense pedestrians,how to improve the representation of pedestrian features during object detection and data association during tracking has remained a difficult topic for MOT in recent years.To solve the above problems,we firstly focus on the object detection stage of the MOT and obtain effective appearance features by employing the CenterNet and enhancing the high-level features,to improve the localization capability of the model.To further improve the tracking performance,the Distance-IoU-based pedestrian multi-object tracking is proposed from the standpoint of data association and combined with object motion information to increase the tracking performance of the model under occlusion.The specific work is as follows:1)A CenterNet-based pedestrian multi-object tracking is proposed.The method can enhance the expressiveness of different feature information and the discriminative ability of the model by aggregating features of different scales;while decoupling multi-branch specific features from the shared feature information to improve the localization and tracking performance of the model.Specifically,the Context Enhancement Module(CEM)is first added to the top-down path of CenterNet to enhance the high-level feature of each layer,followed by the Channel Attention Module(CAM)in each horizontal connection of the network to optimize the expressiveness of the fused features at each level.Finally,the shared head of the network is decoupled into two specific branches for target localization and target classification respectively to further improve the detection and tracking performance.Compared with the Baseline method,the experimental results show that the proposed method improves MOTA and IDF1 increased by 0.1% and 1.0%,respectively,and reduced ID switching by 609 on the MOT17.2)A Distance-IoU-based pedestrian multi-object tracking is proposed.Based on the above study 1),we first employ the strategy of matching between low-score detection and unmatched tracking targets to recover the occluded objects in the low-score detection;Then D istance-IoU is introduced to measure the distance of motion features between objects,which combines the overlapping area and center distance between two bounding boxes to make the model more robust in the case of occlusion;Finally,the occluded objects are processed in blocks,and the local feature of the unoccluded sub-blocks is used for the tracking of the object.Therefore,the proposed improves the object tracking performance in the occlusion and reduces ID switching.Compared with the Baseline method,MOTA and IDF1 increased by0.7% and 2.3%,respectively,and reduced ID switching by 663 on the MOT17.