A new spin-image based 3D Map registration algorithm using low-dimensional feature space

Spin image is a good feature descriptor of the 3D surface, thus it has been extensively used in many applications such as SLAM of mobile robot and cooperation of heterogeneous robots. However, due to the huge computational burden, it is difficult to be used in real time applications. Thus, in order to improve the efficiency and accuracy of spin image based point clouds registration algorithm, a fast registration algorithm is proposed in this paper based on low-dimensional feature space composed of curvature, the Tsallis entropy of the spin image and laser reflection intensity. The main contribution of this paper is that through constructing the low dimensional feature space, the correspondence searching procedure can be divided into two steps: firstly, select similar key points in the proposed low dimensional feature space using k-d tree; then spin image feature is used to search for correspondences among a very limited amount of point candidates. Finally, experiments with respect to a man made surroundings are conducted and the results show the feasibility and validity of the new proposed algorithm. Index Terms - Point cloud map registration, spin image, k-d tree, features description.