立体视觉的微分几何约束与特征匹配

本文在微分几何的统一框架下导出视差，边缘方向和曲率在体视图对上差值或变化的概率分布，在适当的风险下求得相应的匹配界限；视差梯度界限、方向界限和相对曲率界限，从而得到了立体视觉匹配中非常有力的约束条件。据此我们提出一个两步的体视匹配算法，在要素匹配阶段，采用相对较小并与视差相关的匹配界限快速搜索初始匹配；在松驰匹配阶段，形状连续性是唯一的匹配准则，两步匹配过程都借助了投票表决的方法，实验结果表明我们

DIFFERENTIAL GEOMETRY CONSTRAINTS FOR STEREO VISION AND FEATURE MATCHING

Some probability distributions of differential geometry features are derived in this paper. Some stereo matching limits are then obtained by truncating the distributions with proper risks: disparity gradient limit (DGL), orientation limit (OL), and relative curvature limit (RCL). These shape similarity limits provide fairly strong constraints in correspondence. A two-step stereo matching algorithmbased on these limits is proposed. In the intensive matching step, the method uses relatively small, initial disparity related differential shape similarity limits for better candidate match searching. In the relaxation step, figural continuity is the only matching criterion. Both steps are implemented by voting schemes. Experimental results show that our algorithm is robust and efficient.