一种基于分割的可变权值和视差估计的立体匹配算法

Segmentation-Based Stereo Matching Algorithm with Variable Support and Disparity Estimation

立体匹配一直是计算机视觉研究领域中的热点和难点,是立体视觉中的关键技术之一。为了消除基于局部图像的双目立体匹配的歧义性。提出一种基于图像分割及可变权值方案的初始匹配和贪婪的后处理视差估计策略相结合的立体匹配算法。分割彩色立体图像对,利用分割自适应地分配权值来消除匹配特征相似的歧义性,计算匹配代价得到初始视差。接着,为了更好地消除弱纹理区域、重复纹理区域和宽遮挡区域等复杂歧义性,视差后处理中采用贪婪估计方案,包括基于分割的视差校准、窄遮挡处理及多方向自适应加权最小二乘拟合填充。实验结果表明,基于分割的本算法结构简单,能有效地提高处理局外点的稳健性,并生成高精度的稠密视差。

Stereo matching is a long active topic and difficult problem in computer vision, and is a crucial technique in stereovision. An algorithm by combining initial matching via segmentation-based variable support with greedy disparity estimation as post-processing is proposed to resolve the ambiguity of binocular stereo problem in a local perspective. Firstly, color segmentation is conducted on both stereo images, and segmentation-based adaptive support weight is assigned for each pixel to eliminate ambiguity in feature matching, and then matching cost with the variable support is calculated to obtain initial disparity. Secondly, to address more other complex ambiguity in low textured and repetitive patterns or large occluded regions etc., greedy disparity estimation procedure consists sequentially of three steps： segmentation-based disparity calibration, narrow occlusion handling and multi-directional weighted least square fitting. The experimental results indicate that this technique with segmentation cues can increase robustness against outliers and obtain accurate and dense disparity effectively. It＇s concise and efficient.