Registration Method for Low-Overlap Indoor Point Cloud of RGB-D Camera Located by LiDAR and Multirectangle Features

Abstract

As the key to the multiple view sensor fusion of the RGB camera, the point cloud registration (PCR) of the sensor has an important influence on the 3-D reconstruction of the indoor point cloud. A high-precision sensor registration method, including the initial sensor registration and the fine sensor registration, is proposed, to address the issue of the insufficient robustness to low-overlap indoor point clouds and noise in the sensor registration. In the initial sensor registration, the multirectangle feature (MRF) registration for low-overlap point clouds is presented on the basis of the MRF target, the light detection and ranging (LiDAR), and the RGB-D camera. The infinite point and infinite line are generated from the MRF, which is located by the LiDAR. The initial transformation model is constructed by analyzing the MRF. In the fine sensor registration, the point cloud feature extraction method is developed based on the mutual information, the normal vector difference, and the curvature, for the accurate extraction of key feature points. Moreover, a similarity measurement that considers the mutual information, normal vectors, and curvature features of sensor point clouds is explored to achieve the accurate selection of matching sensor point pairs. The average rotation error and average translation error of the method on self-built and public datasets with different overlap rates are 0.008 rad and 0.014 m, respectively. The experimental results on real sensor data and public sensor datasets show that the method enhances the accuracy of the sensor registration of indoor point clouds.