A novel place recognition method for large-scale forest scenes

Abstract

Accurate robot place recognition in expansive forest environments is a crucial challenge. The presence of extensive occlusion areas, repetitive terrain features, and constrained remote communication significantly amplifies the difficulty of place recognition. To address this challenge, we propose an innovative two-stage place recognition approach, termed LGC, which integrates both local and global descriptors. During the coarse matching stage, a trunk contour point cloud segmentation method is developed based on spatial and planar clustering to extract trunk position coordinates. Then, a local triangle descriptor (LTD) is constructed based on these coordinates. This descriptor correlates the structural relationships between local real-world trees and is viewpoint invariant. Next, a hash database is then established, enabling the retrieval of candidate frame sets through a voting mechanism. Finally, geometric consistency verification is performed on the candidate frames. During the fine matching stage, we introduce a Global Trunk Descriptor (GTD) that leverages trunk distribution information to further refine the matching process. This method comprehensively considers the local geometric and global appearance features available in forest scenes, enabling accurate place recognition in large-scale forest scenes. Extensive evaluations are conducted using both publicly available unstructured and real-world forest datasets, benchmarking LGC against other state-of-the-art methods. The results demonstrate that LGC achieves superior adaptability and notable accuracy improvements in forest scenarios, even under challenging conditions with significant viewpoint variations.