针对传统DP算法精度不高, 出现条纹瑕疵以及边界区域明显误匹配等问题, 提出了一种新的立体匹配算法。该算法采用水平、垂直及物体边界多方向约束作为平滑性约束, 物体边界方向约束针对物体边界像素点, 以前一行边界点的视差信息对当前边界点进行约束, 强化物体边界像素点的视差不连续性, 提升了边界区域像素点匹配准确率, 大大减少了边界误匹配现象。由于物体表面的纹理信息会影响物体边界方向约束的准确性, 对物体的边缘提取方法进行了处理, 使提取的物体边缘只保留物体边界轮廓信息。最后在三状态DP的状态转换选择上, 应用边界轮廓信息及多方向约束, 避免了三状态状态转换与实际物理情况不符的局面, 提升了匹配精度。实验表明, 该方法不仅解决了条纹瑕疵、边界区域误匹配问题, 且执行速度快, 匹配精度高。

The performance of conventional DP has not been satisfactory and horizontal streaks occur when applied to the stereo matching problems. To solve these problems, in this paper, we propose a new stereo matching algorithm using the constraints in horizontal, vertical directions and in the direction of the edges of objects as the smoothness. For the boundary pixels of the image, Constraining the current pixels by use of previous matching information to strengthen the object boundary pixels' disparity discontinuities .We also improves the object edge detection method in order to exclude unnecessary edge information. At last we apply the boundary contour information and multi-directional constraints to the choice of state transitions of tri-state DP. The experiment shows that this method alleviate the effect of horizontal steaks, mismatches in the direction of boundary and has a high matching accuracy and a fast speed.