刘杰, 张建勋, 代煜. 基于区域增长的稠密立体匹配[J]. 机器人, 2017, 39(2): 182-188. DOI: 10.13973/j.cnki.robot.2017.0182
引用本文: 刘杰, 张建勋, 代煜. 基于区域增长的稠密立体匹配[J]. 机器人, 2017, 39(2): 182-188. DOI: 10.13973/j.cnki.robot.2017.0182
LIU Jie, ZHANG Jianxun, DAI Yu. Dense Stereo Matching Based on Region Growing[J]. ROBOT, 2017, 39(2): 182-188. DOI: 10.13973/j.cnki.robot.2017.0182
Citation: LIU Jie, ZHANG Jianxun, DAI Yu. Dense Stereo Matching Based on Region Growing[J]. ROBOT, 2017, 39(2): 182-188. DOI: 10.13973/j.cnki.robot.2017.0182

基于区域增长的稠密立体匹配

Dense Stereo Matching Based on Region Growing

  • 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: For dense stereo matching, a local stereo matching algorithm based on region growing is proposed to solve the difficulties in choosing the size and shape of support windows as well as the pixels that have similar depth, and the difficulties in dealing with the boundaries of objects as well as the occlusion areas. Firstly, color similarity and connectivity constraint are adopted to obtain the adaptive support window dynamically by means of a method based on region growing. Then, the weights of pixels in support window or search window are assigned by different strategies, to reduce the effect of the pixels with different depths on the support window, and abandon the support windows that don't conform to the shape similarity rule. Finally, the matching cost is computed. The initial disparity is obtained by means of WTA (winner takes all) strategy. For the support windows, the frequencies of disparities of the pixels in accord with the confidence level standard are computed, in which the disparity with the highest frequency is taken as the optimal disparity. Moreover, the algorithm complexity is reduced through analyzing and optimizing the processes of support region computation, cost aggregation, and disparity search. The proposed algorithm is tested on Middlebury platform, and the result shows its excellent performance.

     

/

返回文章
返回