迭代策略下的行人多目标跟踪研究

Study on Pedestrian Multi-Object Tracking Based on Iterative Strategy

多目标跟踪是计算机视觉领域被广泛研究的重要方向,但在实际应用中,目标的快速移动、光照变化、遮挡等问题会导致跟踪性能变差。在本文中提出了一种迭代策略的行人多目标跟踪方法。采用迭代检测方式,可以在两次迭代分别检测出高置信度和低置信度的行人目标,由于前一次迭代检测到的行人预测框将在下一次迭代中以历史特征的形式传递到网络,从而可以避免模型重复检测同一对象,同时提高行人检测的精度。在数据关联阶段,优先对第一次迭代检测结果进行轨迹匹配即高置信度行人检测框,然后是第二次迭代检测结果,这种对检测结果分批次处理可以有效的减少跟踪过程中身份切换问题。在MOT16数据集的实验表明本文方法对行人目标跟踪的可行性和有效性。

Multi-object tracking is an important research direction in the field of computer vision, but in practical applications, the fast movement of the object, illumination change, occlusion and other problems will lead to poor tracking performance. In this paper, we propose a pedestrian multi-object tracking method based on iterative strategy. By adopting the iterative detection method, pedestrian targets with high confidence and low confidence can be detected respectively in two iterations. Since the pedestrian prediction box detected in the previous iteration will be transmitted to the network in the form of historical features in the next iteration, the model can avoid repeated detection of the same object and improve the accuracy of pedestrian detection. In the data association stage, the trajectory matching of the first iteration detection results, namely, the high confidence pedestrian detection box, is given priority, followed by the second iteration detection results. This batch processing of the detection results can effectively reduce the problem of identity switching in the tracking process. Experiments on MOT16 data sets show that the proposed method is feasible and effective for pedestrian target tracking.