Abstract:

For the problem of feature points dynamic matching in online estimation of object pose, a dynamic matching method of target corners guided by optical flow is proposed, in which the wood block is used as the experimental object. Yolov2-tiny model is used to detect the object for location and classification. Shi-Tomasi algorithm is used to extract corners among the region of proposal. The LK optical flow algorithm is used for corner matching by tracking corners between two adjacent frames. For the problem of feature points migration during optical flow tracking, the corners are re-extracted in the 11×11 neighborhood centered on the tracking points. Experiments show that the algorithm can dynamically match the target corners in the video sequence in real time, and corner matching results are more accurate and real-time than SIFT and Harris-SIFT algorithm.

Key words: Yolov2-tiny, Shi-Tomasi corner, Lucas-Kanade（LK） optical flow, corner match

摘要：

针对目标物位姿在线估计中的特征点动态匹配问题，以木块作为实验对象，提出了一种光流引导的目标物角点动态匹配方法。使用Yolov2-tiny进行目标物识别，完成目标物定位与分类。在提取区域内完成Shi-Tomasi角点检测，采用LK（Lucas-Kanade）光流对角点跟踪的方法实现角点在相邻两帧图像中的初始匹配，针对光流跟踪中特征点的偏移问题，对跟踪点为中心的11×11邻域进行角点再提纯。实验证明该算法能在视频序列中对目标物角点实时动态地匹配，且角点匹配结果比SIFT和Harris-SIFT方法准确度更高，实时性更好。

关键词: Yolov2-tiny, Shi-Tomasi角点, LK光流, 角点匹配