Abstract: Aiming at the accuracy problems in texture lacking region, occlusion region and depth discontinuous in binocular stereo matching, an algorithm based on multi-feature representation and super-pixel optimization is proposed. By adding edge information into initial cost calculating, and combining with image local information, it can improve the edge region recognition in disparity calculation. In cost aggregation step, the initial aggregation region is computed by simple linear iterative clustering method. In order to aggregate much more information in texture lacking region, an algorithm of adaptive searching based on the rice skeleton is proposed. In disparity optimization step, using the initial super-pixel region, to correct disparities which are mismatched. Experiments on the Middlebury stereoscopic dataset test platform prove that the proposed algorithm has higher accuracy.

Key words: binocular stereo matching, multi-feature representation, super-pixel segmentation, super-pixel optimization, computer vision

摘要： 针对目前许多局部双目立体匹配方法在缺乏纹理区域、遮挡区域、深度不连续区域匹配精度低的问题，提出了基于多特征表示和超像素优化的立体匹配算法。通过在代价计算步骤中加入边缘信息特征，与图像局部信息代价相融合，增加了在视差计算时边缘区域的辨识度；在代价聚合步骤，基于超像素分割形成的超像素区域，利用米字骨架自适应搜索，得到聚合区域，对初始代价进行聚合；在视差精化步骤利用超像素分割信息，对匹配错误视差进行修正，提高匹配精度。基于Middlebury立体视觉数据集测试平台，与自适应权重AD-Census、FA等方法得出的视差图进行比较，该算法在深度不连续区域和缺乏纹理区域的匹配效果显著改善，提高了立体匹配精度。

关键词: 双目立体匹配, 多特征表示, 超像素分割, 超像素优化, 机器视觉