Abstract: RGB-D camera, as a common portable 3D data acquisition tool, has been widely utilized in many fields, such as computer graphics, computer vision. However, the captured data is usually corrupted by large noise due to several reasons. For examples, the error of sensor, light interference, etc. The existing point cloud registration methods are incapable of generating desired registration results when processing such data corrupted by large noise, it will directly lead to the quality degradation of the subsequent 3D reconstruction. To tackle the problem, a robust point cloud registration method based on RGB-D images is introduced. First, a point normal estimation method based on a novel second-order operator is proposed, which can effectively remove the noise and simultaneously recover sharp features and nonlinear smooth regions. Then, a compound descriptor, used to robustly extract initial correspondences from noisy point clouds, is designed by combining both texture and geometry information obtained from RGB images and point clouds respectively. Finally, a fast optimization-based method is utilized to calculate the rigid transformation between a pair of point clouds. Intensive experiments on a variety of synthetic and raw point cloud data are conducted, and the experimental results are analyzed from points of the visual effect and the registration error. The analysis results demonstrate the effectiveness of the proposed method, especially in robustly registering point clouds corrupted by large noise. © 2022, Beijing China Science Journal Publishing Co. Ltd. All right reserved.