Abstract: Aiming at the problem that the traditional stereo matching algorithm has low matching accuracy and a single cost function can not deal with strong texture region, weak texture region and non texture region at the same time, a multi-scale stereo matching algorithm based on adaptive texture region is proposed. Firstly, the adaptive threshold is set to divide the image into strong texture region, weak texture region and no texture region. Different scale information fusion is used to calculate the cost. In the strong texture region, AD and census are used to calculate the cost, and in the weak texture region and no texture region, AD and gradient are used to calculate the cost. In the cost aggregation part, in order to solve the problem that the traditional cross arm region construction process can not dynamically adjust the arm length according to different pixels, the gray value constraint is weakened, and the multi-threshold gradient constraint is added to increase the cross arm length of weak texture and non texture regions, so as to significantly optimize the aggregation process of weak texture and non texture regions, that can reduce the false matching rate. Finally, the disparity map is obtained by disparity calculation and multi-step disparity optimization. The experiment is carried out on the Middlebury test platform. The matching accuracy of different cost functions is compared under three conditions：different illumination, different exposure and no distortion, so the multi-scale fusion cost function is proposed. At the same time, compared with the other seven mainstream algorithms, the experimental results show that the proposed algorithm has a final average mismatch rate of 3.57 percentage points lower than ELAS. Compared with LESC, this algorithm has relatively stronger real-time performance. Compared with the single cost function and aggregation stereo matching algorithm, this algorithm has strong adaptability to different texture regions and good robustness.

Key words: machine vision, adaptive texture partition, stereo matching, multi-scale fusion, aggregate cross arm construction

摘要： 针对传统立体匹配算法匹配精度低，且单一代价函数不能同时应对强纹理区域、弱纹理和无纹理区域的问题，提出一种自适应纹理区域的多尺度融合立体匹配算法。首先设置自适应阈值，将图像划分为强纹理区域、弱纹理和无纹理区域，采用不同的尺度信息融合计算代价，在强纹理区域采用AD与Census相结合计算代价，在弱纹理和无纹理区域采用AD与梯度相结合计算代价；在代价聚合部分，为了解决传统十字臂区域构建过程不能根据不同像素点动态调整臂长的问题，弱化灰度值约束条件，同时添加多阈值梯度约束条件，增加弱纹理和无纹理区域的十字臂长，显著优化弱纹理和无纹理区域聚合过程，达到降低误匹配率的目的；最终通过视差计算和多步视差优化得到视差图。在Middlebury测试平台上进行实验，选择不同光照、不同曝光和无失真三种情况，对比不同代价函数的匹配精度，从而提出多尺度融合代价函数；同时和其他七种主流算法对比代价聚合后的匹配精度，实验结果表明，所提算法较于ELAS，最终平均误匹配率下降3.57个百分点；相较于LESC，实时性相对更强。相比单一的代价函数和聚合方式的立体匹配算法，此算法对不同纹理区域的适应性强，具有很好的鲁棒性。

关键词: 机器视觉, 自适应纹理划分, 立体匹配, 多尺度融合, 聚合十字臂构建