基于多特征SAD-Census变换的立体匹配算法

Stereo matching algorithm based on multi-feature SAD-Census transformation

视差不连续区域和重复纹理区域的误匹配率高一直是影响双目立体匹配测量精度的主要问题,为此,本文提出一种基于多特征融合的立体匹配算法。首先,在代价计算阶段,通过高斯加权法赋予邻域像素点的权值,从而优化绝对差之和(Sum of Absolute Differences,SAD)算法的计算精度。接着,基于Census变换改进二进制链码方式,将邻域内像素的平均灰度值与梯度图像的灰度均值相融合,进而建立左右图像对应点的判断依据并优化其编码长度。然后,构建基于十字交叉法与改进的引导滤波器相融合的聚合方法,从而实现视差值再分配,以降低误匹配率。最后,通过赢家通吃(Winner Take All,WTA)算法获取初始视差,并采用左右一致性检测方法及亚像素法提高匹配精度,从而获取最终的视差结果。实验结果表明,在Middlebury数据集的测试中,所提SAD-Census算法的平均非遮挡区域和全部区域的误匹配率为分别为2.67%和5.69%,测量200~900 mm距离的平均误差小于2%;而实际三维测量的最大误差为1.5%。实验结果检验了所提算法的有效性和可靠性。

The high mismatching rate of the parallax discontinuity region and the repeated texture region has been a major issue affecting the measurement accuracy of binocular stereo matching.For these reasons,we propose a stereo matching algorithm based on multi-feature fusion.Firstly,the weight of neighboring pixels is given using Gaussian weighting method,which optimizes the calculation accuracy of the Sum of Absolute Differences(SAD)algorithm.Based on the Census transformation,the binary chain code technique has been enhanced to fuse the average gray value of neighborhood pixels with the average gray value of gradient image,and then the judgment basis of the left and right image corresponding points is established,and the cod-ing length is optimized.Secondly,an aggregation technique has been developed that combines the cross method and the improved guide filter to redistribute disparity values with the aim of minimizing false match-ing rate.Finally,the initial disparity is obtained by the Winner Take All(WTA)algorithm,and the final dis-parity results are obtained by the left-right consistency detection method,sub-pixel method,and then a stereo matching algorithm based on the multi-feature SAD Census transform is established.The experimental res-ults show that in the testing of the Middlebury dataset,the average mismatch rates of the proposed algorithm for non-occluded regions and all regions are 2.67%and 5.69%,the average error of the 200−900 mm dis-tance is less than 2%,and the maximum error of the actual 3D data measurement is 1.5%.Experimental res-ults verify the effectiveness of the proposed algorithm.