Diaz-Ramirez Victor H, Gonzalez-Ruiz Martin, Juarez-Salazar Rigoberto, Cazorla Miguel
Instituto Politécnico Nacional-CITEDI, Ave. Instituto Politécnico Nacional 1310, Tijuana 22310, Mexico.
CONAHCYT, Instituto Politécnico Nacional-CITEDI, Ave. Instituto Politécnico Nacional 1310, Tijuana 22310, Mexico.
Sensors (Basel). 2024 Dec 24;25(1):21. doi: 10.3390/s25010021.
Accurate estimation of three-dimensional (3D) information from captured images is essential in numerous computer vision applications. Although binocular stereo vision has been extensively investigated for this task, its reliability is conditioned by the baseline between cameras. A larger baseline improves the resolution of disparity estimation but increases the probability of matching errors. This research presents a reliable method for disparity estimation through progressive baseline increases in multiocular vision. First, a robust rectification method for multiocular images is introduced, satisfying epipolar constraints and minimizing induced distortion. This method can improve rectification error by 25% for binocular images and 80% for multiocular images compared to well-known existing methods. Next, a dense disparity map is estimated by stereo matching from the rectified images with the shortest baseline. Afterwards, the disparity map for the subsequent images with an extended baseline is estimated within a short optimized interval, minimizing the probability of matching errors and further error propagation. This process is iterated until the disparity map for the images with the longest baseline is obtained. The proposed method increases disparity estimation accuracy by 20% for multiocular images compared to a similar existing method. The proposed approach enables accurate scene characterization and spatial point computation from disparity maps with improved resolution. The effectiveness of the proposed method is verified through exhaustive evaluations using well-known multiocular image datasets and physical scenes, achieving superior performance over similar existing methods in terms of objective measures.
在众多计算机视觉应用中，从捕获的图像中准确估计三维（3D）信息至关重要。尽管双目立体视觉已针对此任务进行了广泛研究，但其可靠性受相机之间基线的制约。较大的基线可提高视差估计的分辨率，但会增加匹配错误的概率。本研究提出了一种通过在多目视觉中逐步增加基线来进行视差估计的可靠方法。首先，引入了一种针对多目图像的鲁棒校正方法，该方法满足极线约束并最小化诱导失真。与现有的知名方法相比，该方法可将双目图像的校正误差降低25%，将多目图像的校正误差降低80%。接下来，通过立体匹配从具有最短基线的校正图像中估计密集视差图。然后，在短优化区间内估计具有扩展基线的后续图像的视差图，将匹配错误和进一步误差传播的概率降至最低。此过程反复进行，直到获得具有最长基线的图像的视差图。与类似的现有方法相比，所提出的方法可将多目图像的视差估计精度提高20%。所提出方法能够通过具有更高分辨率的视差图进行准确的场景表征和空间点计算。通过使用知名的多目图像数据集和物理场景进行详尽评估，验证了所提出方法的有效性，在客观指标方面比类似的现有方法具有更优的性能。
