Abstract: Multi-object tracking can be divided into two sub-tasks： target detection and data association. With the increasing complexity of multi-object tracking datasets, the problem of occlusion caused by dense crowds significantly affects the accuracy of detection and association. To address this, two innovative modules are proposed. One is detection repair module, it repairs the results of target detection. The highest response point obtained by cross-correlating the appearance features of the previous frame trajectory with the current frame features serves as a potential detection box. It is fused with the initial detection box to achieve detection restoration. The concept of local feature expansion is introduced to mitigate false positives generated by this approach. Another is smooth feature update module,how to determine and measure the degree of occlusion is a prerequisite for further optimization. For this reason, feature independence is proposed as a measure of occlusion. The lower boundary of the detection frame is used to determine the occlusion relationship. The overlapped part of the detection frame is compared to itself to calculate the feature independence. On this basis, a smooth appearance feature update algorithm is designed. On the dense multi-object tracking dataset MOT20, state-of-the-art performance is achieved on a single-stage model,with HOTA improving to 55.3% and MOTA improving to 70.7%.

Key words: multi-object tracking, repair detection, occlusion, appearance update

摘要： 多目标跟踪可以分为两个子任务，目标检测任务和数据关联任务。随着多目标跟踪数据集的日益复杂，密集人群引起的遮挡问题严重影响了检测和关联的准确性。为此提出了两个改进模块。检测修复模块，对目标检测的结果进行修复。前一帧轨迹的外观特征和当前帧特征作互相关运算得到的最高相应点作为潜在的检测框，将其与初始检测框融合，实现检测修复，并引入特征局部扩展思想减少该方式产生的误检。平滑的特征更新模块，如何确定和衡量遮挡程度是进一步优化的前提。为此提出特征独立性作为衡量遮挡指标，利用检测框下边界判断遮挡关系，以检测框重合部分比上自身计算特征独立性，并在此基础上设计平滑的外观特征更新算法。在密集多目标跟踪数据集MOT20上，用单阶段模型实现了最先进的性能，HOTA提升到55.3%和MOTA提升到70.7%。

关键词: 多目标跟踪, 检测修复, 遮挡, 外观特征更新