Abstract: Multi-object tracking (MOT) aims to identify and track multiple targets in the video sequence, and maintain the identity of each target. Tracking multiple targets with irregular movements is a difficulty in this field. Moreover, the tracking accuracy is difficult to guarantee and the target ID is prone to frequent switching. In order to improve the tracking accuracy, the paper proposes the adaptive Kalman filter (AKF) to make full use of the information provided by the target detector to modify the motion model and improve the tracking accuracy of the moving irregular targets. In order to solve the problem of frequent switching of target ID, a greedy algorithm BIoUG is designed to increase the matching chance and reduce the mismatching probability by enlarging the matching frame and adopting the preferred matching method. On this basis, a multi-target tracking method for moving irregular targets is proposed. The experimental results show that the MOTA, HOTA and IDF1 of the proposed method on DanceTrack dataset reach 92.2%, 57.7% and 58.7%, respectively; on MOT17 dataset, MOTA, HOTA and IDF1 reach 80.3%, 63.3% and 77.3%, respectively. Compared with the current mainstream target tracking methods, the proposed method has better tracking effect on both irregular and regular targets, embodies better comprehensive performance, and provides a new way to solve the problem of irregular multi-target tracking.

Key words: multi-object tracking(MOT), Kalman filter, data association, trajectory matching

摘要： 多目标跟踪MOT（multi-object tracking）旨在对视频序列中的多个目标进行识别与跟踪，并保持各目标的ID（identity），对运动不规律的多目标进行跟踪是该领域的难点，跟踪准确度难以保证且易出现目标ID频繁切换的问题。为提高跟踪准确度，提出自适应卡尔曼滤波AKF（adaptive Kalman filter），充分利用目标检测器提供的信息对运动模型加以修正，提高对运动不规律目标跟踪准确度；为解决目标ID频繁切换的问题，设计一种BIoUG贪婪算法，通过放大匹配框并采取择优匹配的方式，提高匹配机会并降低误匹配概率。在此基础上，提出一种针对运动不规律目标的多目标跟踪方法。实验结果表明，所提方法在DanceTrack数据集上的MOTA、HOTA、IDF1分别达到了92.2%、57.7%和58.7%；在MOT17数据集上，MOTA、HOTA、IDF1分别达到了80.3%、63.3%和77.3%。与目前主流的同类目标跟踪方法相比，所提方法对运动不规律和规律的目标均有较好的跟踪效果，体现出较好的综合性能，为运动不规律多目标的跟踪提供了新的解决思路。

关键词: 多目标跟踪, 卡尔曼滤波, 数据关联, 轨迹匹配