LiDAR-Based Place Recognition For Autonomous Driving: A Survey

TL;DR

This survey delivers the first comprehensive, LiDAR-focused treatment of place recognition for autonomous driving, defining explicit problem formulations for loop closure and global localization and detailing seven subdivisions across handcrafted and learning-based approaches. It consolidates a wide range of techniques—local and global descriptors, segment- and semantics-based methods, trajectory- and map-assisted strategies—while assessing datasets, metrics, and practical evaluation results. The work highlights critical challenges such as motion distortion, viewpoint shifts, and long-term environmental changes, and it emphasizes the growing role of multimodal and advanced-sensor integrations. By providing a detailed taxonomy, pragmatic observations, and an up-to-date public project, the paper aims to accelerate the development and standardized evaluation of robust LPR methods for real-world autonomous driving deployments.

Abstract

LiDAR-based place recognition (LPR) plays a pivotal role in autonomous driving, which assists Simultaneous Localization and Mapping (SLAM) systems in reducing accumulated errors and achieving reliable localization. However, existing reviews predominantly concentrate on visual place recognition (VPR) methods. Despite the recent remarkable progress in LPR, to the best of our knowledge, there is no dedicated systematic review in this area. This paper bridges the gap by providing a comprehensive review of place recognition methods employing LiDAR sensors, thus facilitating and encouraging further research. We commence by delving into the problem formulation of place recognition, exploring existing challenges, and describing relations to previous surveys. Subsequently, we conduct an in-depth review of related research, which offers detailed classifications, strengths and weaknesses, and architectures. Finally, we summarize existing datasets, commonly used evaluation metrics, and comprehensive evaluation results from various methods on public datasets. This paper can serve as a valuable tutorial for newcomers entering the field of place recognition and for researchers interested in long-term robot localization. We pledge to maintain an up-to-date project on our website https://github.com/ShiPC-AI/LPR-Survey.

Figures (9)

• Figure 1: Two key problems addressed by PR. On the left, the blue line denotes the vehicle trajectory, with solid circles indicating collected scans by sensors over time. The green circle marks the current scan, while the others represent past scans. The green and red circles are geographically close, and their scans exhibit the highest scene similarity, forming a closed loop. On the right, the black-marked area shows a vehicle's global location, only providing a single-location description within maps.
• Figure 2: The relationship between PR and key components of autonomous driving. HD maps, being external information, are grouped with sensors under data modules.
• Figure 3: Proposed taxonomy of LPR methods.
• Figure 4: Definition of LiDAR-based place recognition. We classify the definition into implicit loop closure detection and explicit global localization. Since global localization provides global vehicle poses, we term it explicit place recognition. In contrast, loop closure detection identifies revisited poses through frame-to-frame comparison without a global pose, making it implicit.
• Figure 5: Two projection-based methods. (a) and (b) are originally shown in 2022-RAL-OverlapTransformer and 2021-RAL-DiSCO, respectively.