基于改进Census变换的鲁棒立体匹配算法  

Robust Stereo Matching Algorithm Based on Improved Census Transform

作　　者：乔景慧[1] 韩玉明 张啸涵 QIAO Jing-hui;HAN Yu-ming;ZHANG Xiao-han(College of Mechanical Engineering,Shenyang University of Technology,Shenyang,Liaoning 110870,China)

机构地区：[1]沈阳工业大学机械工程学院,辽宁沈阳110870

出　　处：《计量学报》2023年第5期694-700,共7页Acta Metrologica Sinica

基　　金：国家自然科学基金(61573249);辽宁省自然科学基金(2019-MS-246);辽宁省教育厅基金(LZGD2019002);辽宁省高等学校创新人才项目(LR2019048);辽宁省研究生教育教学改革研究资助项目(LNYJG2022073);沈阳工业大学重点科研基金(ZDZRGD2020004)。

摘　　要：针对局部立体匹配算法对噪声与光照变化敏感及在弱纹理区域匹配效果不佳的情况,提出一种基于改进Census变换与自适应参数引导滤波的立体匹配算法。该算法融合HSV通道值计算代价,对噪声与光照变化有较好的鲁棒性;在支持窗口内分别使用欧式距离加权与颜色加权获得Census变换参考值;集成AD与Census代价提高了单像素匹配代价的稳定性,降低了原Census变换对中心像素的依赖程度。在代价聚合过程中使用峰度系数对正则化参数进行自适应处理,通过视差计算获得视差图。在VS2017平台上对Middlebury数据库提供的图像对进行匹配实验,该算法在标准图像、加噪声图像及光照变化图像的各区域平均误匹配率分别是7.80%、10.72%和9.89%。结果表明:该算法可以降低误匹配率,同时能更好地适应噪声与光照变化。Aiming at the problems of the noise and illumination sensitive and with high false matching radio in the weak texture region of the local matching algorithm,an robust stereo matching algorithm for improved Census transform and adaptive parameter-guided filter is proposed.The hue,saturation,value(HSV)channel value is used to calculate the cost,which has higher robustness to noise and illumination changes.In the support window,the distance weighting and color weighting are used to obtain the reference value of the census transform respectively.The two costs are fused to improve the stability of the single pixel matching cost.The kurtosis adaptive regularization parameter is used on the guide filter for cost aggregation.Finally,the disparity map is obtained by disparity calculation.The matching experiment is carried out on the image pairs provided by Middlebury database on the VS2017.The average error matching rate of the proposed algorithm of standard image,noisy image and illumination change image is 7.80%,10.72%and 9.89%respectively.The results show that the algorithm can reduce the false matching ratio and better adapt to the changes of noise and illumination.

关 键 词：计量学 立体匹配 图像处理 改进Census变换 代价计算 自适应参数 误匹配率 

分 类 号：TB96[机械工程—光学工程]