基于区域增长的稠密立体匹配

发布时间：2019-02-09 13:16

【摘要】：针对稠密立体匹配中支持窗口的尺寸、形状和深度相似像素难以选择,以及物体边界和遮挡区域难以处理的问题,提出一种基于区域增长的局部立体匹配算法.首先应用区域增长的方法,通过颜色相似性和连通性约束动态地获得完全自适应的支持窗口.然后运用不同的策略为支持窗口内像素和搜索窗口内其他像素分配权值,减弱非深度相似像素对支持窗口的影响,并摒弃视差范围内不符合形状相似标准的支持窗口.最后,计算匹配代价.应用WTA(赢者全取)策略获得初始视差,统计支持窗口内符合置信度条件的像素的视差频率,以出现频率最高的视差作为最优视差.通过对支持区域获取、代价聚合和视差搜索步骤的分析和优化,降低算法复杂度.通过Middlebury平台的检验,表明算法具有良好的性能.
[Abstract]:A local stereo matching algorithm based on region growth is proposed to solve the problem that the size, shape and depth of the supporting window are difficult to select in dense stereo matching, and the object boundary and occlusion region are difficult to deal with. Firstly, a fully adaptive support window is obtained dynamically by means of color similarity and connectivity constraints using the method of region growth. Then different strategies are used to support the distribution of weights between the pixels in the window and the other pixels in the search window to reduce the influence of the non-depth similar pixels on the support window and to abandon the support window which does not conform to the shape similarity standard in the parallax range. Finally, the matching cost is calculated. The initial parallax is obtained by using the WTA (winner-take-all) strategy, and the parallax frequency of pixels in the support window that meets the confidence condition is calculated. The parallax with the highest frequency is used as the optimal parallax. The complexity of the algorithm is reduced by analyzing and optimizing the steps of support region acquisition, cost aggregation and parallax search. The test of Middlebury platform shows that the algorithm has good performance.
【作者单位】：南开大学机器人与信息自动化研究所;
【基金】：国家自然科学基金(51375494) 天津市应用基础与前沿技术研究计划(14JCQNJC04600)
【分类号】：TP391.41

【相似文献】