基于视差区域分割的动态规划立体匹配算法改进

Improvements of Dynamic Programming Stereo Matching Algorithm Based on Parallax Region Segmentation

基于动态规划的立体匹配算法在较低的硬件条件下,也可以满足实时性的要求,因此,可以在基于立体视觉的机器人导航避障系统中应用。但传统动态规划算法存在匹配精度不高、易出现分散畸变点等问题,因此,论文对动态规划算法初始匹配代价求取、路径寻径及回溯等加以改进。在初始代价求取阶段,提出了一种变窗口能量聚集法,通过获取场景的视差变化区域与视差连续区域的位置信息,从而使像素点在能量聚合时能够根据视差变化自适应地调整聚合窗口的大小,使能量聚合方式更加合理,提高了初始视差的准确性;在路径寻径及回溯阶段,使用多路径寻径回溯法,保留更多的可靠点,减少了误匹配现象的发生。因此,提高了立体匹配的匹配精度,并具有较好的实时性。

Stereo matching algorithm based on dynamic programming （DP） can meet real-time constraints even on low-cost hardware. Therefore, it can be used in Robot obstacle avoidancing system. But the performance of conventional DP has not been satisfactory and scattered mismatching points occur when the stereo matching is applied. To solve these problems, the method of initial cost getting and path traversal and backtracking is improved. The energy aggregation method is proposed which can use the information of object boundary to determine the disparity changing regions. A highly accurate initial disparity map is obtained, which can make the subsequent phases of parallax getting get a good performance. In the path traversal and backtracking phase, the idea of multi-path backtracking to exploit the information gained from DP more effectively is also introduced. More reliable pixels to reduce the occurrence of mismatch can be retained. The experiment shows that this method has a high matching accuracy and a fast speed.