基于图像片段匹配的立体匹配算法

An Stereo Matching Method Based on Fragment Matching of Image

摘要:

提出一种以图像片段为单位进行立体匹配的方法，与传统的以像素为单位的立体方法不同的是，本文使用图像的边缘信息将参考图像分成小的片段，同时在水平和垂直方向上以小的片段为单位进行立体匹配，然后进行视差图像的融合，提高了匹配的准确性.在片段内，本文将匹配代价分为高代价与低代价，以不同的计算方式统计两种代价值对整个片段的代价贡献，这种计算方式可以准确地匹配物体被部分遮挡时的情况；接下来对原参考图像进行低位数放大，对片段内的视差值根据低位数放大后的参考图像再次查找精确视差；最后，依据原始的参考图像，使用同色域均值滤波对获得的视差图进行填补与平滑操作.经实验证明，本方法具有算法运算复杂度低，速度快，匹配准确和物体的边缘视差清晰的特点，在Middlebury图像库的测试中获得了较好的实验结果.

Abstract:

A method of stereo matching using each image fragments as a unit is proposed. Different from the conventional pixelbypixel stereo method， the authors use the edge information of the image to divide the reference image into small segments and perform stereo matching in units of small segments in the horizontal and vertical directions， and then use the disparity images to improve the accuracy of the matching. In each segment， the authors divide the matching cost into high cost and low cost and calculate the cost contribution of the two cost values to the entire segment in different calculation modes. This calculation method can accurately match the situation when the object is partially occluded.Next， the authors enlarge the loworder digit of the intensity for the two reference images and search for more accurate disparity values in all the fragments according to the enlarged reference images. Finally， the disparity map is filled and smoothed using the same color region median filter according to the original reference image.The results of the experiments show that this method has the characteristics of low computational complexity， fast speed， accurate matching， and clear edge disparity of the object. It achieved good performance with the Middlebury stereo benchmark.