结合多路梯度与重排序的AD-Census立体匹配算法

AD-Census stereo matching algorithm combining multiple gradients and reordering

针对局部立体匹配算法在深度不连续和弱纹理区域匹配精度低,且容易受到噪声干扰的问题,提出一种结合多路梯度与重排序的AD-Census立体匹配算法。对传统的AD-Census变换进行改进,重排序不同尺度变换窗口中的像素,取中值作为中心像素点计算Hamming距离,使其对噪声具有更强的鲁棒性,并将八方向梯度信息相融合进行代价计算以增强可靠性;采用十字交叉域和4路径扫描线优化进行代价聚合,提高匹配精度;采用赢家通吃法则进行视差计算,同时引入多步骤视差优化策略得到最终的视差图。在Middlebury测试平台提供的标准立体图像上,所提算法的平均误匹配率低至4.62%,与经典和新颖算法相比,匹配精度明显提高,具有更强的稳健性。

Stereo matching technology is one of the key technologies in the field of binocular vision,and it is also a focus in current research.It calculates the disparity of matching pixel pairs in binocular images to accurately estimate the depth of objects in a 3D scene.However,in practice,due to the influence of various complex factors such as lens distortion,lighting variation,and occlusion,it has become a challenging task to accurately determine whether the content of two images matches.Although the stereo matching technology has made significant progress and a relatively complete technical system has been established,there are still some problems that need to be solved urgently.Among them,the traditional AD(Absolute Difference)-Census transformation method is a prominent problem case.This method is based on a fixed matching window and relies too much on the center pixel of the window,so when the center pixel is affected by unfavorable factors such as lighting,its value is prone to abrupt changes,which will interfere with the accuracy of the AD-Census transformation value.This interference not only increases the risk of mismatching the corresponding pixels in the disparity map,but also has a serious impact on the stability of the algorithm.Therefore,the existing stereo matching algorithms based on AD-Census have obvious limitations in obtaining the final disparity results,which urgently requires the development of a new algorithm that is more stable and robust.In order to overcome the above problems,this paper proposes an innovative matching algorithm,namely the AD-Census algorithm combining multi-path gradient and reordering.Firstly,the traditional AD-Census transform is refined by reordering pixels within varying scale transformation windows and using the median value as the center pixel for Hamming distance calculations.This modification improves robustness to noise and incorporates eight-direction gradient information for reliable cost calculation.Furthermore,cost aggregation is optimized through the utilization of cross-shaped support regions and four-path scan line optimizations,elevating matching precision.Finally,the Winner-Take-All(WTA)strategy is adopted for disparity calculation,complemented by a multi-step disparity optimization approach to yield the ultimate disparity image.The new algorithm inherits the traditional AD-Census method based on the window matching characteristics and ranks the pixels in the matching window by weighting them,and replaces the central pixels that are abruptly affected by noise by taking the median value.In doing so,the matching relationship between pixels can be more accurately reflected,and the interference caused by the abrupt change of the center pixel in traditional algorithms can be effectively reduced.In addition,the eight-direction gradient information further improves the matching accuracy of the algorithm in the depth discontinuous region and weak texture region at the edge of the image.In order to verify the effectiveness of the new algorithm,this paper conducts extensive experimental verification on the standard stereo image pair on the Middlebury test platform.The results show that compared with the classical and novel stereo matching algorithms,the proposed algorithm achieves a significant reduction in the false matching rate,reaching 4.62%.This result fully proves the excellent performance of the new algorithm in terms of accuracy.At the same time,the new algorithm also shows obvious advantages in terms of running time,which further proves its high efficiency.In general,the AD-Census algorithm combining multi-path gradient and reordering proposed in this paper not only successfully overcomes the limitations of traditional algorithms,but also significantly improves the matching accuracy by introducing gradient information.This innovative algorithm has injected new vitality into the development of stereo matching technology,which is expected to promote its wide application in many fields such as 3D reconstruction and robot navigation.