多特征融合的匹配代价与视差优化算法研究

Matching Costs and Disparity Refinement Algorithm Based on Multi-features Fusion

针对现有算法对光照变化敏感、非连续区域与纹理区域易出现误匹配的问题,本文提出了一种多特征融合的匹配代价与视差优化算法.该方法首先分两阶段计算匹配代价,第一阶段联合使用三种特征:颜色特征、WCCT特征、LBP特征.基于上述三种特征的互补性,通过加权融合计算初始匹配代价;第二阶段使用SIFT特征代价修正初始匹配代价.然后,使用改进最小生成树聚合匹配代价与初始视差图计算.最后,使用置信度聚合与传播策略进行视差图优化,得到高质量的视差图.实验结果表明,在多种复杂场景中,该算法都能够提高立体匹配精度.

To address some challenges in stereo matching,e. g. sensitiveness to light changes and mismatch in texture and boundary regions.Matching costs and disparity refinement algorithm based on multi-features fusion is proposed. Our method calculates the matching costs in two stages,Firstly,multiple features are used： color feature,WCCT feature,and LBP feature. The three kinds of features are combined to calculate the initial matching costs,which are then aggregated on the improved minimum spanning tree. Secondly,using SIFT features amended the initial matching costs. Finally,belief aggregation and propagation is used for refining the disparity map. Experimental results demonstrate that the proposed method can improve the accuracy of stereo matching.