A 3D Sequential LiDAR Data Registration Method for Unmanned Ground Vehicle

Abstract

This paper proposes a 3D point cloud registration method based on light detection and ranging (LiDAR) system. The proposed method consists of three steps: Gaussian-Process based ground segmentation, a novel k-neighbors based dynamic point feature and Iterative Closest Point (ICP) fine registration. The first two steps are the preparation of ICP fine registration. The odometry information from a GPS/IMU system is used to compensate the vehicle's ego-motion. The Gaussian-Process based ground segmentation is adopted to remove ground points. A novel Initial Localization based Dynamic Feature (ILDF) is proposed to detect and remove dynamic points. It is applicable in sequential frames and a proper initial localization without a large dislocation. In experiment results, a large number of dynamic points will be detected and removed by ILDF. The removal of dynamic points improves both accuracy and efficiency of registration algorithm.