排除光流错误跟踪点的鲁棒方法

摘要

Optical flow is a widely used method for points tracking. However, the performance of optical flow decreases drastically when the illumination changes or partly occlusion happens. To solve this problem, we present a novel method for outlier rejections from optical flow tracking points. By measuring the distance between the source point and the destination point obtained from its complements' projection relation, our method automatically determines the point with largest error. We prove the correctness of our method and then give the camera pose matrix on video sequence. The experimental comparison of our results to those of random sample consensus (RANSAC) shows the robustness of our method.%针对光流法用于跟踪光照变化和部分遮挡情况下的物体容易产生漂移的问题,通过比较跟踪点与该点补集映射关系产生的投影点之间的距离,提出了一种光流错误跟踪点排除方法——异类距法.首先证明了异类距排除误差最大元素的正确性；然后在静止场景受光照变化和部分遮挡情况下,给出了异类距排除错误跟踪点以及摄像机姿态矩阵在时间序列上的分布.针对视频序列的抖动情况,与传统方法进行比较的实验结果表明,该方法对于排除错误跟踪点是鲁棒的.