针对立体匹配中弱纹理区域和深度不连续区域的匹配精度问题, 提出了一种基于多特征融合的树形结构代价聚合立体匹配算法. 首先, 融合图像颜色、梯度和图像的 Census 变换进行匹配代价计算; 然后, 在由原始图像生成的最小生成树上进行匹配代价聚合, 并使用多方向扫描线优化, 进一步提升立体匹配的精确度; 最后, 使用左右一致性检测标记出误匹配点, 并进行视差修正. 为了验证该算法的有效性, 使用 Middlebury 测试集提供的测试图像进行测试, 平均误匹配率为 6.38%; 分别对 2 种场景实际拍摄图像进行深度信息提取误差率测试, 测试得到 2 种场景的测距误差率分别为 5.76% 和 5.55%, 证明了该算法的实用性.

reo matching in texture and depth discontinuities, a stereo matching algorithm based on multi-feature fusion and tree structure aggregation was proposed. The first step was color image fusion, gradient image Census transform and cost calculation. Next came the matching cost aggregation in the minimum spanning tree generated by the original image. Then the accuracy of stereo matching was further improved through using multi direction scanning line optimization. Finally, through using the left and right consistency check, error matching points were marked and parallax correction is made. In order to verify the effectiveness of the proposed algorithm, and the first set of test images provided were tested using the Middlebury test, and the average error rate is 6.38%; then on both actual scene images error rate testing for depth information extracted was conducted, and the test scenarios point to error rates at 5.76% and 5.55% respectively, demonstrating the practicality of the algorithm.