基于特征匹配的非刚性大位移光流算法

Non-Rigid and Large Displacement Optical Flow Based on Descriptor Matching

针对非刚性大位移光流估计问题,提出了一种基于特征匹配的光流优化算法.首先,传统数据项针对整副图像进行一致性假设的策略过于粗放,文中采用了一种根据图像中每个像素特点自适应选择一致性假设的优化方法;其次,针对传统各向同性的平滑项易造成演化过程中图像细节丢失、边缘模糊等问题,提出一种各向异性的平滑项,有利于实现图像细节的光流估计;另外融合特征匹配于变分光流框架中,充分利用特征匹配在大位移条件下的鲁棒性与变分光流的致密性;最后,基于MPI Sintel数据集对本文算法进行定量分析.实验结果表明,本文算法能够实现非刚性大位移光流的准确估计,鲁棒性强,在market_5图像中,该算法比LDOF于AAE和AEPE指标上分别提升了18.9%与21.9%.

In order to solve the arithmetic problem of non-rigid and large displacement optical flow,an improved arithmetic was proposed based on descriptor matching for optical flow.In the method,a novel data term was put forward firstly to adaptively select invariable hypothesizes for each pixel.And then,the traditional isotropic smoothness term was optimized to be the anisotropic one for the optical flow estimation of image details.Furthermore,the descriptor matching was applied to optical flow fields,taking advantage of the robust ability of descriptor matching to produce some large displacement correspondences and the ability of dense optical flow.At last,a quantitative analysis of the approach was performed on MPI Sintel.The results show that the proposed method can realize accurate estimation of non-rigid and large displacement optical flow.And the method is superior to LDOF with a relative gain of 18.9%for AAE and 21.9%for AEPE in market_5.