半全局立体匹配算法的改进研究

Improvement of Semi-global Stereo Matching Algorithm

针对半全局立体匹配中单一Census精度不足且边缘区域视差效果较差的问题,提出一种基于颜色差信息BTCensus和加入分割约束的半全局匹配算法。首先该算法采用联合三通道BT算法和Census变换的计算方式实现代价计算,缓解重复纹理歧义并提高不连续区域匹配精度;同时将图像分割得到的场景信息作为约束对代价聚合函数进行改进,保证割块内部的纹理平滑,提高边界纹理区域的匹配精度;最后引入图像分割信息进行分步中值滤波,平滑视差图。实验表明,该算法在重复纹理区域、视差不连续区域和弱纹理区域都取得了较好的视差效果,可获得高匹配率的视差图。

A semi-global matching algorithm based on color difference information BTCensus and segmentation constraints is proposed to solve the problem of poor precision of single Census and poor disparity in edge region.Firstly, the algorithm uses the combined three-channel BT algorithm and Census transform to realize the cost calculation, which can reduce the repeated texture ambiguity and improve the matching precision of discontinuous regions.At the same time, the scene information obtained from image segmentation is used as a constraint to improve the cost aggregation function to guarantee the texture smoothness and improve the matching precision.Finally, image segmentation information is introduced to filter the median step and smooth the disparity map.The experimental results show that the algorithm has good disparity effect in repeat texture area, disparity discontinuity area and weak texture area, and can get disparity map with high matching rate.