An improved feature extractor for the Lidar Odometry and Mapping (LOAM) algorithm

Abstract

This work proposes an improved feature extractor for the Lidar Odometry and Mapping (LOAM) algorithm, which is currently the highest ranked algorithm in the Karlsruhe Institute of Technology and Toyota Technological Institute (KITTI) visual odometry ranking. This article proposes and justifies the substitution of LOAM's current feature extraction method with the Curvature Scale Space (CSS) based feature extraction algorithm for the processing of 3D Point Cloud Data (PCD). The justification is based on in improvement of the repeatability of the detection of robust features for LOAM and an improvement in the associated computational cost. The LOAM's feature extractor and CSS feature extractor were tested and compared with simulated and real data including the KITTI visual odometry dataset using the Optimal Sub-Pattern Assignment (OSPA) and Absolute Trajectory Error (ATE) metrics. The results showed that LOAM based on the CSS feature extractor out performed that based on the original LOAM feature extractor with respect to these metrics.