光照变化条件下结合图像纹理与改进census变换的鲁棒光流算法

针对传统光流算法只能对某种特定光照变化有效的问题,提出一种能改善不同光照变化条件下的光流鲁棒求解方法.所提出的方法利用改进census变换和图像纹理信息加权和构建光流模型中数据项.数据项中不同分量的权重根据两帧图像SIFT点对应灰度差动态选择,以满足不同亮度变化的图像,自适应地选择相应的数据项.改进census变换包含了变换窗口内完整的相对灰度大小信息;对两个不同的窗口,较传统census变换有更高的区分度.以Middlebury和KITTI数据库对所控算法进行测试,实验结果证实了所提出方法对不同仿真光照变化和野外真实光照环境下图像鲁棒光流求解的有效性.

Using Image Texture and Improved Census Transformation to Improve the Robustness of Variational Optical Flow Against Illumination Changes

The traditional optical flow algorithm can only perform well under a special variation illumination condition. In order to solve this problem, an algorithm which can improve the optical flow computation robustness under different illumination changes was proposed. An improved census transformation and image texture were used to construct data items in the optical flow model. The weightings of different components in the data term are dynamically determined according to the grayscale differences of SIFT points of the two frames, which can lead that the corresponding data term can be chosen in different illumination conditions. The improved census transformation includes completed relative gray information in the transformation window which guarantees it has better discrimination for two different transformation windows. Middlebury and KITTI databases were used to test the proposed algorithm. Experimental results verify the effectiveness of the robust optical flow calculation of the proposed method for real outdoor images.