A Closed-Form Solution for Coarse Registration of Point Clouds Using Linear Features

Abstract

This paper presents a closed-form procedure for the coarse registration of three-dimensional (3D) point clouds using automatically extracted linear features, which have been manually matched. Corresponding linear features are defined by nonconjugate endpoints that do not necessarily define compatible direction vectors. Because the point clouds could be derived from different sources (e.g., laser scanning data sets and/or photogrammetric point clouds that are referenced to arbitrary reference frames), the proposed procedure estimates the scale, shift, and rotation parameters that relate the reference frames of these data sets. The proposed approach starts with a quaternion-based procedure for initial estimation of the transformation parameters using the minimal number of required conjugate line pairs (i.e., two noncoplanar linear features from each point cloud). The initial estimate of the transformation parameters is then used to ensure the compatibility of the direction vectors of the involved linear features. The modified direction vectors together with the endpoints of the linear features are used for deriving a better estimate of the transformation parameters. Experimental results from both simulated and real data sets verified the feasibility of the proposed procedure in providing good quality for the approximate parameters of the transformation parameters for point-based fine registration procedures.