基于改进代价计算和自适应引导滤波的立体匹配  

Stereo Matching Method Based on Improved Cost Computation and Adaptive Guided Filter

作　　者：闫利[1] 王芮 刘华[1] 陈长军[1] Yan Li;Wang Rui;Liu Hua;Chen Changjun(School of Geodesy and Geomatics,Wuhan University,Wuhan,Hubei 430079,China)

机构地区：[1]武汉大学测绘学院,湖北武汉430079

出　　处：《光学学报》2018年第11期249-259,共11页Acta Optica Sinica

基　　金：国家重点研发计划(2017YFC0803802)

摘　　要：针对现有局部立体匹配算法在弱纹理区域匹配精度低的问题,提出一种基于改进代价计算和自适应引导滤波代价聚合的局部立体匹配算法。该算法首先将增强后的梯度信息与基于增强梯度的Census变换相结合,构建代价计算函数;然后对图像的每一个像素构建自适应形状十字交叉窗口,并基于自适应窗口进行引导滤波代价聚合;最后通过视差计算和多步视差精化得到最终的视差图。实验结果表明,改进后的算法在Middlebury测试平台上对标准立体图像对的平均误匹配率为4.80%,与基于传统引导滤波器的立体匹配算法相比,本文算法在弱纹理区域取得更好的匹配结果。To solve the problem of low matching accuracy in textureless regions, a local stereo matching method is proposed based on improved cost computation and adaptive shape guided filter. First, an efficient cost function combining enhanced image gradient and enhanced gradient-based Census transform is introduced for cost computation. Then, an adaptive shape cross-based window is constructed for each pixel, and guided filter aggregation is implemented based on this adaptive window. The final disparity map is obtained after disparity computation and multi-step disparity refinement. The experimental results demonstrate that the average matching error rate of the proposed algorithm is 4.80 % for stardard image paris on Middlebury benchmark. Compared with traditional guided filter-based method, the proposed method has better matching results in textureless regions.

关 键 词：机器视觉 立体匹配 代价计算 自适应引导滤波 Census变换 

分 类 号：P232[天文地球—摄影测量与遥感]