Natural Terrain Detection and SLAM Using LIDAR for an UGV

Abstract

This paper describes natural terrain detection and SLAM algorithms using a LIDAR sensor for an unmanned ground vehicle. We describe how features are detected from natural terrain, and then localize the vehicle’s position and compose a map with the detected features. The LIDAR is equipped on the experimental vehicle to scan natural terrain. The scan data include many kinds of natural disturbances on uneven terrain: a banded tree, a branch of a tree, non-uniform size of bushes, and undefined or unexpected objects. We apply a RANSAC (RANdom SAmple Consensus) algorithm to discriminate ground point cloud data and object point cloud data, and then separate bush points and tree points by a combination of two algorithms, GMM (Gaussian Mixture Model) and EM (Expectation Maximization). The GMM and EM algorithms are for extracting features and classifying groups, respectively. We propose a double FCM (Fuzzy C-mean clustering) algorithm to robustly estimate the number of trees and their positions. The Extended Kalman Filter applied to simultaneous localization and mapping (EKF-SLAM) is composed of the extracted tree features. The mahalanobis distance is applied to retain consistency for feature correspondence, which is used data association. Finally, we show the results obtained from implementation of the approach on a tree-covered mountain.