光学学报, 2017, 37 (11): 1115004, 网络出版: 2018-09-07
基于改进Census变换和异常值剔除的抗噪立体匹配算法 下载: 796次
Anti-Noise Stereo Matching Algorithm Based on Improved Census Transform and Outlier Elimination
机器视觉 立体匹配 抗噪 Census变换 双目测量 machine vision stereo matching anti-noise census transform binocular measurement
摘要
针对Census变换易受噪声影响使得立体匹配算法难以获取高匹配精度的问题,提出了一种改进Census变换和异常值剔除的抗噪立体匹配算法。在初始匹配代价阶段,该方法首先将窗口邻域中值作为参考值并通过映射函数控制异常值,提高了单像素匹配代价的可靠性;然后在代价聚合阶段,对动态聚合窗口中初始代价值进行异常值剔除;最后通过视差计算、视差优化得到最终的视差图。在VS2013软件平台上采用Middlebury标准测试图对初始匹配代价、代价聚合、最终视差图阶段进行测试。实验结果表明,本文算法的抗噪性能优于现有Census变换算法,且错误匹配率达到5.71%。
Abstract
The Census transform is sensitive to noise, so it is difficult to obtain high matching accuracy with stereo matching algorithm. An anti-noise stereo matching algorithm based on the improved Census transform and outlier elimination is proposed. Firstly, at the initial match cost stage, the median of the window neighborhood is taken as the reference value and the outliers are controlled by the mapping function, which improves the reliability of the single pixel matching cost. At the cost aggregation stage, the outliers are eliminated from the initial cost value of dynamic aggregation window. Finally, the final disparity maps are obtained by disparity calculation and disparity optimization. The Middlebury benchmark images are used to test the stages of the initial matching cost, cost aggregation, and final disparity map on the VS2013 software platform. Experimental results show that the proposed algorithm has better noise-robust performance than the existing Census transform algorithms, and the error matching rate is 5.71%.
彭新俊, 韩军, 汤踊, 尚裕之, 俞玉瑾. 基于改进Census变换和异常值剔除的抗噪立体匹配算法[J]. 光学学报, 2017, 37(11): 1115004. Xinjun Peng, Jun Han, Yong Tang, Yuzhi Shang, Yujin Yu. Anti-Noise Stereo Matching Algorithm Based on Improved Census Transform and Outlier Elimination[J]. Acta Optica Sinica, 2017, 37(11): 1115004.