부산대학교 도서관

Geo-localisation in GPS-poor environments such as forests is crucial in field robotics and remains a challenge. To tackle this problem, we introduce a collaborative localisation framework that fuses ‘above canopy’ height information obtained from airborne aggregated lidar scans, as a reference map, with information of trees sensed under canopy using a 3D lidar sensor on a mobile platform. Under the canopy we extract information, i.e., position of trees and their crown height, invariant to both the ground and the overhead viewpoints, generating an ‘under canopy’ local map. We then compare the above canopy reference map with the under canopy local map using a similarity score to localise the robot within the reference map. We use a Monte Carlo localisation algorithm to incorporate the similarity between maps by defining a set of particles to track the robot pose hypotheses and converge to the true solution using the robot motion. Experimental evaluation on three different platforms over different forest scenarios validates the presented method. Our approach achieved sub-meter average position error, demonstrating its effectiveness to geo-localise ground robots in dense foliage.