一种基于光流的行人阴影检测与跟踪

Pedestrian Shadow Detection and Tracking Based on Optical Flow

提出一种基于光流和机器学习的行人阴影检测与跟踪算法。算法假设初始帧的阴影边缘已知,首先利用从已知结果中提取的边缘信息特征对RF(随机森林)模型进行训练,然后在对前后两帧进行光流跟踪的同时找出误差较大的点,然后利用训练好的RF模型从邻域Canny获选边缘点中识别符合阴影特征的点。由于视点移动可能会带来的有新场景进入画面的问题,采用动态更新RF的方法。最后,对于仍可能存在的阴影边缘断裂不连续或者错误分类的情况,可以使用一些方法进行补全和修正。实验结果表明,该算法可以准确地检测和跟踪移动视点下视频中运动人体的投射阴影。

Proposes a pedestrian shadow detection and tracking algorithm based on optical flow and machine learning. The algorithm assumes that the shadow edges of the initial frame are known. Firstly, the RF (random forest) model is trained by using the edge information features extracted from the known results. Then, the points with large errors are found while tracking the optical flow of the two frames. Then, the trained RF model is used to identify the points with shadow features from the selected edge points of neighborhood Canny. Because of the problem that new scenes may enter the screen due to viewpoint movement, we adopt the method of dynamic updating RF. Finally, some methods can be used to complete and correct the discontinuous or wrong classification of shadow edge breaks that may still exist. The experimental results show that the algorithm can accurately detect and track the projected shadow of moving human body in video from moving viewpoint.