一种针对大尺度运动的快速光流算法

传统的光流法必须满足一致连贯性假设,不适用于大幅度运动目标的跟踪。为克服复杂场景下,大尺度运动中目标位移量超出特征跟踪窗口带来的光流计算问题,提出了一种针对大尺度运动的快速光流计算方法。引入图像金字塔模型,利用基于多尺度分层的金字塔结构和光流映射变换,实现小窗口捕获大运动。同时,采用非极大值抑制方法保留图像自相关矩阵的大特征值,得到的强角点作为特征点,降低了光流计算的时间开销。实验结果表明,提出的方法在复杂场景和运动条件下,可准确快速地计算出表征目标运动方向和速率的光流矢量,具有较高的鲁棒性和实时性。

Fast optical flow algorithm for large-scale motion

Conventional optical flow computation method must satisfy the consistent and coherent hypothesis, which is not appropriate to large-scale motion. In order to overcome the difficulty of calculating optical flow on fast and great magnitudes motion whose target displacement exceeds feature tracking window, this paper introduced the multi-scale image pyramid model and vector mapping transform, adjusting the image resolution to capture large-scale motion within small calculating window. Meanwhile, by using non-maximum suppression method to select points with large eigenvalue of image correlation matrix as strong corner feature, it reduced the cost of optical flow calculation. Experimental results show that the proposed algorithm is robust to complicated circumstances and can obtain the direction and speed of optical flow vector efficiently.