Matched Filtering based LiDAR Place Recognition for Urban and Natural Environments

TL;DR

The paper addresses the challenge of robust LiDAR place recognition across diverse environments with minimal training. It introduces a two-stage, roto-translation invariant LPR pipeline based on low- and high-resolution BEV descriptors and a matched-filter search, enabling direct relative pose estimation. The approach yields state-of-the-art performance on both urban and natural datasets (NCLT, Oxford Radar, WildPlaces), with notable recall gains and accurate pose estimates, while maintaining practical runtimes. This work advances generalization of LPR methods to unstructured environments and offers a promising plug-in for SLAM-based relocalisation and loop closure.

Abstract

Place recognition is an important task within autonomous navigation, involving the re-identification of previously visited locations from an initial traverse. Unlike visual place recognition (VPR), LiDAR place recognition (LPR) is tolerant to changes in lighting, seasons, and textures, leading to high performance on benchmark datasets from structured urban environments. However, there is a growing need for methods that can operate in diverse environments with high performance and minimal training. In this paper, we propose a handcrafted matching strategy that performs roto-translation invariant place recognition and relative pose estimation for both urban and unstructured natural environments. Our approach constructs Birds Eye View (BEV) global descriptors and employs a two-stage search using matched filtering -- a signal processing technique for detecting known signals amidst noise. Extensive testing on the NCLT, Oxford Radar, and WildPlaces datasets consistently demonstrates state-of-the-art (SoTA) performance across place recognition and relative pose estimation metrics, with up to 15% higher recall than previous SoTA.

Figures (7)

• Figure 1: Matched filter-based LiDAR Place Recognition (LPR) architecture. Low-resolution reference BEV descriptors are accumulated within a grid while the query BEV descriptor is rotated for a global search using the matched filter. The top $n$ matches from this search are then used for a high-resolution local search. The resulting correlation output from the top match is used for translation estimation and pose correction from reference to query traverse.
• Figure 2: Illustrated BEV descriptor generation process with a scan from the WildPlaces Dataset knights2023wild: a.) Point cloud data cropped by a 3D window and voxelised. b.) BEV image created from a height density map. c.) Thresholded height density map forming a BEV descriptor indicating occupancy. d.) BEV descriptor after randomly downsampling patches. e.) Lower resolution BEV descriptor after average pooling.
• Figure 3: Sample LiDAR scans from three different datasets: 1.) NCLT Dataset scan using Velodyne 32E-HDL LiDAR sensor captured in an urban driving environment from Michigan carlevaris2016university. 2.) WildPlaces Dataset scan using Velodyne 16-VPL LiDAR sensor captured in unstructured natural environments from Brisbane knights2023wild. 3.) Oxford Radar Dataset scan using Velodyne32-VPL LiDAR sensor captured in an urban driving environment from Oxford barnes2020oxford.
• Figure 4: Long Term Place Recognition Performance. Examination of LPR performance for query collected from day one, tested on references from the same day, 6 months later and 14 months later using 5 different methods, including ours.
• Figure 5: Impact of the relative shift on position estimation for the complete WildPlaces dataset: The position error density distribution of correct LPR matches with and without relative pose correction demonstrates a lower mean error and smaller deviation after correction.