基于单目视觉的多车辆跟踪方法研究

Research on Tracking Method of Multiple Vehicles Based on Monocular Camera

针对车载单目摄像机的汽车辅助驾驶系统,提出了一种多模式多线索的车辆跟踪算法。首先利用运动估计预测目标在下一帧的感兴趣区域,也就是跟踪窗。然后多目标遮挡推理机制确定目标是处于遮挡模式还是正常模式。在正常模式下,在跟踪窗内联合使用车辆的先验知识、评估环节、分块模板等线索对车辆目标定位;遮挡模式下,利用遮挡推理机制给出的遮挡区域,在跟踪窗内调用切割分块模板匹配进行遮挡处理。这种多模式多线索的车辆跟踪方法实现了对后方多车辆的稳定跟踪,同时也适合前视车辆。实验结果表明该算法是鲁棒的和准确的。

The multi-mode and multi-clue.vehicle tracking algorithm in image sequences was proposed for automobile assistance driving system with monocular camera. The regions of interest which contain the target in the next frame, namely tracking windows, are predicted by motion estimator. Either the occlusive mode or common mode of the target is determined by the multi-target occlusion reasoning mechanism. In the common mode, multiple clues, such as prior knowledge of vehicle, assessment component, division block template etc, are integrated to locate the vehicle target in the tracking window; in the occlusive mode, partial division block template matching is used to handle with occlusion. The vehicle observed in rear view can be tracked stably by the multi-mode and multi-clue vehicle tracking algorithm. The approach is suitable for tracking vehicles observed in frontal view. Experiment resuits illustrate the robust and accurate performance of the tracking algorithm.