摘要
为解决传统最小生成树(MST)立体匹配算法在纹理重复和深度不连续区域的匹配精度较低问题,提出一种基于双目视觉融合多维信息的最小生成树立体匹配算法。在原始代价函数中引入垂直梯度和像素截断绝对差(truncated absolute difference, TAD)共同衡量;融合像素的颜色距离和颜色内相关性决定树结构的边权值,以降低相似颜色处的像素点连接错误问题,构建最小生成树进行代价聚合。改进算法在重复纹理和深度不连续区域得到了较好的匹配效果,匹配误差率降低到5.02%。
To address the low matching accuracy of the traditional minimum spanning tree(MST)stereo matching algorithm in the areas with texture duplication and depth discontinuities,this paper proposes an improved MST stereo matching algorithm based on binocular vision fusion multi-dimensional information.In the original cost function,vertical gradient and truncated absolute difference(TAD)are introduced as additional measures.The edge weight of the tree structure is determined by fusing the color distance and intra-color correlation of pixels,reducing errors in pixels connections of similar colors.The improved algorithm constructs a minimum spanning tree for cost aggregation and achieves better matching results in repetitive texture and depth-discontinuous areas,reducing the matching error rate to 5.02%.
作者
周斌
宫刘莉
黄丰云
ZHOU Bin;GONG Liuli;HUANG Fengyun(School of Mechanical and Electronic Engineering,Wuhan University of Technology,Wuhan 430070,China)
基金
工业和信息化部工业互联网创新发展工程资助项目(TC19084DY)
关键词
最小生成树
垂直梯度
颜色内相关性
多维权重
代价聚合
minimum spanning tree
vertical gradient
intra-color correlation
multi-dimensional weight
cost aggregation
作者简介
周斌(1976-),男,湖北武汉人,武汉理工大学机电工程学院副教授,博士.
