Multispectral Registration Algorithm Based on Fusion of Infrared and Color Images

Multispectral data fusion is a fundamental research in the fields of artificial intelligence and computer vision, which provides technical support and theoretical basis for remote sensing, autonomous driving, and medical imaging. In this paper, we present a robust registration algorithm to align RGB-D scans. In colored point clouds, limited by the principle of point cloud collection and generation, noise, and outliers are common and may cause registration failure. Therefore, the core of our method is to apply color information and kernel function to remove the influence of noise and outliers on the alignment accuracy. In the beginning, for the first time, we innovatively use the Cauchy kernel as the energy function of the point set registration to assign different weights to inliers, noise, and outliers to maximize their differentiation. Combined with the maximization optimization strategy, noise and outliers are removed to achieve robust registration. Secondly, since RGB-D data contain both shape and color information, the trimmed bidirectional color distance is applied to reduce the probability of noise and outliers being selected as the matching point pairs when establishing the correspondences. Moreover, the overall formulation is solved by the iterative optimization method. Experiments demonstrate that our algorithm is more robust and accurate than previous point set registration algorithms, especially when aligning point sets containing noise and outliers.