Abstract: To vanquish the problems of low accuracy and mismatching in weak texture region,an Anisotropic Stereo Matching algorithm with Multi-Scale Information(ASMSI)was proposed.An anisotropic matching cost function was constructed,the gradient and phase angle information were introduced into the cost calculation process to eliminate outliers in the weak texture region.Besides,cross-based cost aggregation was adopted to calculate the matching cost value of each support region.Additionally,a winner-take-all strategy was employed to generate the initial disparity map.On this basis,the refined parallax chart was produced by left and right consistency detection and post-processing.Numerical simulations were carried out to compare the mismatching rate,running time of non-occlusion and depth-discontinuous regions of images to evaluate the algorithm model.The simulation results verified that the presented approach solved the mismatching problems of weak texture region effectively and improve the matching accuracy by 5.02%,which could satisfy the requirements of efficacy and high precision in the stereo matching process.

Key words: machine vision, binocular vision, stereo matching algorithm, cost aggregation, disparity optimization

摘要： 针对传统立体匹配算法准确率低且在弱纹理区域存在误匹配的问题,提出融合多尺度信息的各向异性立体匹配算法(ASMSI)。首先构造各向异性的匹配代价计算函数,将梯度和相角信息引入代价计算过程中用于剔除弱纹理区域的离群点;随后采用融合多尺度信息十字交叉代价聚合计算每个支持域内的匹配代价;进一步经赢家通吃策略生成初始视差图;在此基础上进行左右一致性检测及后处理得到精修后的视差图;最后通过仿真实验对比图像中非遮挡、深度不连续区域的误匹配率和运行时间来评价算法模型。实验结果表明:所提算法能有效解决弱纹理区域的误匹配问题,使匹配准确率提高了5.02%,能够满足立体匹配过程中高效率、高精度的要求。

关键词: 机器视觉, 双目视觉, 立体匹配算法, 代价聚合, 视差优化