Tracking pedestrians based on visual sensors has many diverse applications, among them autonomous driving. Besides obtaining high recall, maintaining the consistency of tracked trajectories during data association is one of the most crucial issues of any tracker. This issue has been tackled in the literature for some time, taking advantage of geometry cues for improving the pairwise matching of detections across consecutive frames. However, this idea has only been employed in a simple way and not thoroughly leveraged in existing studies, i.e., only 2D information is utilized that cannot help to completely understand the real-world geometry in 3D space. Motivated by this observation, in this paper, we present a new method called 3D-TLSR (3D pedestrian tracking using local structure refinement). We use stereo images and expand the idea of geometry cues into 3D space to improve the association of existing trajectories and new detections. We divide the assignment optimization into two steps: (1) determining trajectories whose assignments are strongly believed to be correct, which we call anchors and (2) employing geometry constraints between the anchors and their nearby trajectories in 3D space to improve the matching of less reliable assignments of the first step. In addition, we suggest a simple approach to compute and correct the velocity of a tracked person so that we can better recover missed detections. Experimental results on the well known KITTI tracking benchmark, the ETHMS data set, as well as a self-generated dataset show that our tracker yields comparable results to other state-of-the-art methods with (for KITTI) multi object tracking accuracy (MOTA) of 54.00, which is the best online result among all investigated approaches, multi object tracking precision (MOTP) of 73.03, which is the best of all reported values, and mostly tracked (MT) of 29.55, being the second-best result. On the ETHMS dataset, our approach obtains best results with large margins for recall, precision, and MT, while maintaining a reasonable low number of Id switches (IDs) and fragmentation (FG). These findings confirms the effectiveness of our proposed association method and velocity estimation approach.

基于视觉传感器跟踪行人具有多种应用，其中包括自动驾驶。除了获得较高的召回率外，在数据关联期间保持跟踪轨迹的一致性是任何跟踪器中最关键的问题之一。在文献中已经解决了这个问题一段时间，利用几何线索来改善跨连续帧的检测的成对匹配。但是，这种想法仅以简单的方式被采用，而在现有研究中并未得到充分利用，即仅利用了2D信息，无法完全理解3D空间中的实际几何形状。基于这一观察，本文提出了一种称为3D-TLSR（使用局部结构细化的3D行人跟踪）的新方法。我们使用立体图像并将几何线索的概念扩展到3D空间中，以改善现有轨迹与新检测之间的关联。我们将分配优化分为两个步骤：（1）确定其分配被认为正确的轨迹，我们将其称为锚点；（2）在3D空间中在锚点及其附近的轨迹之间采用几何约束，以改善3D空间中的匹配。第一步的可靠分配。此外，我们建议一种简单的方法来计算和校正被跟踪人员的速度，以便我们可以更好地恢复错过的检测。在著名的KITTI跟踪基准，ETHMS数据集上的实验结果，以及自生成的数据集表明，我们的跟踪器产生的结果与其他最新方法（对于KITTI）的多对象跟踪精度（MOTA）为54.00相当，这是所有研究方法中最佳的在线结果，多目标跟踪精度（MOTP）为73.03，这是所有报告值中最好的，而大多数跟踪（MT）为29.55，是第二好的结果。在ETHMS数据集上，我们的方法获得了最佳结果，并具有较大的查全率，精度和MT余量，同时保持了合理的ID切换（ID）和碎片（FG）数量较低。这些发现证实了我们提出的关联方法和速度估计方法的有效性。这是所有报告值中最好的，并且追踪的（MT）为29.55，是第二好的结果。在ETHMS数据集上，我们的方法获得了最佳的结果，并具有较大的查全率，精度和MT余量，同时保持了合理的ID切换（ID）和分段（FG）数量低。这些发现证实了我们提出的关联方法和速度估计方法的有效性。这是所有报告值中最好的，并且追踪的（MT）为29.55，是第二好的结果。在ETHMS数据集上，我们的方法获得了最佳结果，并具有较大的查全率，精度和MT余量，同时保持了合理的ID切换（ID）和碎片（FG）数量较低。这些发现证实了我们提出的关联方法和速度估计方法的有效性。