分区域快速高精度立体匹配方法

A fast high-precision subregion stereo matching method

针对传统高精度立体匹配策略时间复杂度较高的问题,提出了一种分区域快速立体匹配方法。该方法首先建立加密的Delaunay三角网并计算网点视差。将结合梯度和灰度信息的零均值归一化互相关函数ZNCC作为匹配度量,采用视差邻域区间的二分法对加密三角网的顶点视差值进行亚像素级精化。在加密三角网中,利用视差边缘区域内视差值变化幅度较大的特性,提取出三角网中的边缘三角形。并连接各边缘三角形中视差最大的角点以获得视差边缘,同时对视差边缘进行邻域膨胀得到视差传播区域,其余区域视为视差平滑区域。在全局匹配区域内,选取待匹配点固定范围内代价最小的网点视差作为视差结果,由此获得全局的初始视差值。最后,利用初始视差值对视差传播区域内的像素构造初始视差平面,并进行平面传播和法向量优化以获取区域内的最优视差。结合视差传播区域和平滑区域的视差值,最终得到全局的视差结果;通过对Middlebury dataset中的图像实验对比分析,得出该方法在nonocc和disc指标平均值方面都具有更高的匹配精度,实现了快速的高精度匹配。

Aiming at the high time complexity of the traditional high-precision stereo matching strategy,a sub-regional fast stereo matching method is proposed.The method firstly establishes the encrypted Delaunay triangulation and calculates of vertex disparity.The ZNCC(Zero-normalized cross-correlation)method combined with gradient and grayscale information is used as the matching metric,and the sub-pixel level refinement of the vertex disparity value of the encrypted triangulation is carried out by the dichotomy of the disparity neighborhood interval.In the encrypted triangulation network,the edge triangles in the triangulation network are extracted by using the feature that the disparity in the edge area of the disparity changes greatly.The corner point with the largest disparity in each edge triangle is connected to obtain the disparity edge.At the same time,the disparity edge is neighborhood expanded to obtain the disparity propagation area,and the other areas are regarded as the disparity smoothing area.In the global matching region,the pixel disparity with the lowest cost within the fixed range of the points to be matched is selected as the disparity result to obtain the global initial disparity value.Finally,the initial disparity plane is constructed for the pixels in the disparity propagation area by using the initial disparity value,and the plane propagation and normal vector optimization are performed to obtain the optimal disparity in the area.The global disparity result is finally obtained by combining the disparity values of the disparity propagation region and the smooth region;Through a large number of experiments on images in Middlebury dataset,the method in this paper and PMS(Patch Match),GEC(Global Edge Constraint),CSM(Consistent StereoMatching)and SGF(Stereo Gait Feature),the method in this paper has higher matching accuracy in terms of the average value of nonocc and disc indicators.In particular,the disc indicator is 1.24%lower than the second place,and the matching time is 78.9%lower than PMS,and 29.95%lower than SGF,thus realizing fast and high-precision matching.