Abstract: Multi-Object Tracking (MOT) in the field of computer vision faces numerous challenges, such as target occlusion and appearance similarity, which significantly reduce tracking accuracy and robustness. To address these issues effectively, this paper proposes a new MOT method called TBSTrack. This method involves three core modules: temporal prediction, feature extraction, and stagewise matching. The temporal prediction module constructs a temporal information buffer and uses a self-attention mechanism to calculate the predicted results for the current frame, thereby enhancing the spatiotemporal association of targets and accurately predicting their positions. The feature extraction module handles occluded targets through patch-based processing. It employs a Convolutional Neural Network (CNN) to extract features from each patch and then merges them based on occlusion status, thus effectively eliminating interference and enabling robust target feature representation. The stagewise matching module adopts a two-stage matching strategy, in which learnable anchors are used to recover missed targets during matching and mining of potential targets from the background. The final tracking results are obtained by integrating the recovered and newly mined targets, and the temporal information updated accordingly. Experiments conducted using the MOT17, DanceTrack, and SportsMOT datasets evaluate the performance of TBSTrack. The results show that TBSTrack achieves High Order Tracking Accuracy （HOTA） of 63.9%, 57.3%, and 75.6%, and Identify F1 value （IDF1） of 79.6%, 56.7%, and 78.8%, respectively, on the three datasets. These experimental results demonstrate that TBSTrack significantly improves the accuracy and robustness of MOT, especially in complex scenarios.

Key words: Multi-Object Tracking (MOT), self-attention mechanism, feature extraction, occlusion handling, Convolutional Neural Network (CNN)

摘要： 多目标跟踪(MOT)在计算机视觉领域面临目标遮挡、相似外观等诸多挑战，这些问题极大地制约了跟踪的准确性与鲁棒性。为有效应对此类难题，提出一种新的多目标跟踪方法TBSTrack。该方法由时序预测、特征提取和分段匹配3个核心模块组成。时序预测模块通过构建时序信息缓存区，并结合自注意力机制计算当前帧的预测结果，强化目标的时空关联性，从而精准预测目标位置。特征提取模块针对遮挡目标进行分块处理，运用卷积神经网络（CNN）提取各分块信息，再依据遮挡状况进行拼合，有效去除干扰，实现目标特征的有效表征。分段匹配模块采用两阶段匹配策略，借助可学习锚点在匹配时恢复遗漏目标，再从背景中挖掘潜在目标，综合两者得到最终跟踪结果，以此更新时序信息。为验证方法性能，在MOT17、DanceTrack和SportsMOT数据集上展开实验，结果显示，该方法的高阶跟踪精度（HOTA）分别为63.9%、57.3%、75.6%，识别F1值（IDF1）分别为 79.6%、 56.7%、78.8%。实验结果表明，该方法显著提升了多目标跟踪的准确性和鲁棒性，尤其在复杂场景下优势明显，为多目标跟踪提供了切实有效的解决方案。

关键词: 多目标跟踪, 自注意力机制, 特征提取, 遮挡处理, 卷积神经网络

CLC Number: 

• TP181