基于引导图像和自适应支持域的立体匹配  

Stereo Matching Based on Guidance Image and Adaptive Support Region

作　　者：孔令寅 朱江平 应三丛[1] Kong Lingyin;Zhu Jiangping;Ying Sancong(College of Computer Science,Sichuan University,Chengdu,Sichuan 610065,China)

机构地区：[1]四川大学计算机学院,四川成都610065

出　　处：《光学学报》2020年第9期80-92,共13页Acta Optica Sinica

基　　金：国家自然科学青年基金(61901287);四川省重大科技专项(2018GZDZX0024)。

摘　　要：提出一种基于引导图像和自适应支持域的局部立体匹配算法。首先对校正后的输入图像进行预处理得到引导图像;在匹配代价计算阶段,提出一种梯度计算方法,结合引导图像和输入图像的梯度信息,分别计算x和y方向的梯度,再与AD(absolute difference)和Census变换融合构建匹配代价计算函数;在代价聚合阶段,使用基于自适应支持域的导向滤波;在视差细化阶段,提出一套基于自适应支持域的多步细化方法,通过该方法得到最终的视差图。实验结果表明,视差细化后全部区域的平均误差和方均根误差平均减少43.7%和38%,非遮挡区域平均减少33.7%和30.9%,所提算法具有较好的鲁棒性并能获得精度较高的视差结果。In this study,we propose a local stereo matching algorithm based on guidance images and an adaptive support region.First,the guidance images can be obtained by preprocessing the rectified input images.During the matching cost calculation stage,we propose a gradient calculation method,which combines the gradient information of the guidance and input images to calculate the gradients along the x and y directions,respectively,and subsequently integrates the absolute difference(AD)and the Census transform to develop a matching cost calculation function.Further,we use a guided filter based on the adaptive support region during the cost aggregation stage.During the disparity refinement stage,a multi-step refinement method is proposed based on the adaptive support region and then the final disparity map is obtained.The experimental results prove that after disparity refinement,the average error(Avgerr)and root-mean-square error(RMSE)are reduced by 43.7%and 38%respectively for all the regions and by 33.7%and 30.9%for the non-occluded regions.The proposed algorithm exhibits improved robustness and can be used to obtain high precision disparity results.

关 键 词：机器视觉 局部立体匹配算法 引导图像 自适应支持域 导向滤波 

分 类 号：TP301.6[自动化与计算机技术—计算机系统结构]