SingleMatch: a point cloud coarse registration method with single match point and deep-learning describer

Point cloud registration is a necessary step of object digitization. In this paper, we propose a coarse registration method with a single match point. To achieve the purpose, feature points with stable orientation are recognized firstly, then descriptors of these points are generated with our Convolution Neural Network (CNN) named PFNet. Finally, candidate solutions are obtained by descriptors matching and the accurate registration is given by a RANSAC-based optimization strategy. As the feature points used are highly directional, a stable Local Coordinate System (LCS) can be constructed by combining the orientation and the normal vector, and thus, the registration can be realized by LCS mapping with single match point. Experiment results show that our algorithm achieves good registration effects in challenging scenes, and is robust to noise, outliers, non-uniform sampling.