Visual-LiDAR Localization and Tree Map Building for Autonomous Robot in Unstructured Forests

Abstract

This study proposes a Visual-LiDAR (VL) localization and tree trunk mapping method for an autonomous robot in unstructured forests. In unstructured outdoor environments such as forests and orchards, the GPS signal is often unreliable and occluded by trees. Thus, new Simultaneous Localization and Mapping (SLAM) methods should be investigated rather than conventional SLAM methods developed for structured environments. We employ vision and LiDAR sensors to obtain the image and shape of tree trucks. To use the trees as the SLAM landmarks, we detect trees from the LiDAR scan data and save their trunk models in a tree database. While updating the tree database using sequences of LiDAR scan data, a tree map is also updated and used as the landmarks for LiDAR SLAM. The LiDAR odometry and the stereo-based visual odometry are loosely coupled to enhance the robot localization performance. Experimental results show that the proposed method can reconstruct tree maps in 0.202m average accuracy and complete loop closure in 0.373m average accuracy in unstructured forests. The proposed method is also tested in an arranged orchard testbed, and tree mapping results show only about 0.05m error.