Laser Scan Matching in Polar Coordinates Using Gaussian Process

Abstract

Laser scan matching serves an important role as pose estimation in autonomous navigation for mobile robots. As a classic approach, iterative closest point (ICP) finds corresponding point pairs in a brute-force way, which is time consuming. For eliminating the cost of searching the correspondence between the point pairs, polar scan matching (PSM) uses a matching bearing rule by making use of the laser measurements collected in polar coordinates. However, PSM might occur mismatching when the irrelevant region of the reference and current scans has a similar distribution on polar range. In order to obtain better results, we propose a novel laser scan matching method based on a new type of map representation using GP regression in polar coordinates. With this map representation, the corresponding point pairs can be found in a simple and efficient way. Based on these pairs, we get the robot pose by iteratively estimating the orientation and translation. For dense-point-scan data sets, our approach demonstrates an outstanding performance compared with traditional approaches like ICP and PSM in terms of both accuracy and efficiency.