Abstract
LiDAR-based Place Recognition (LPR) is essential for autonomous vehicle localization and loop closure detection. Traditional multi-view fusion methods often suffer from "information islands," where different views are processed independently until late fusion, and rely on Euclidean metrics that fail to capture complex data distributions. This paper presents a novel cross-view network featuring a Pseudo-Global Fusion Paradigm and a learned Mahalanobis distance metric (MAPVLM). By integrating early contextual guidance through a Pseudo-Global View and warping the feature space for better discriminative power, the proposed method achieves state-of-the-art performance across multiple benchmarks, including NCLT, KITTI, and Ford Campus.
Introduction
LiDAR-based Place Recognition (LPR) serves as the primary mechanism for autonomous vehicles to determine their location by matching current LiDAR scans against a database of previously visited places. Most contemporary LPR methods project 3D point clouds into 2D representations, such as Bird’s Eye View (BEV) and Range View (RV), to leverage established 2D feature extraction techniques.
Despite their success, these methods face two significant challenges. First, the "Information Island" problem: BEV and RV branches are typically processed in isolation, preventing the model from exploiting complementary geometric cues during early feature extraction. Second, the "Euclidean Trap": the use of standard Euclidean distance for metric learning assumes a uniform feature space, which is inadequate for the complex, non-linear distributions and intra-class variances found in real-world LiDAR data.
Methodology
The proposed framework addresses these bottlenecks through two primary innovations: the Pseudo-Global Fusion Paradigm and the MAPVLM metric.
Pseudo-Global Fusion Paradigm
Unlike traditional late-fusion architectures, this paradigm introduces a Pseudo-Global View (PGV) early in the pipeline. The PGV is constructed by concatenating multilayer BEV and RV representations. It serves as a mediator that provides global contextual guidance to individual view branches via a Contextual Alignment Module (CAM). By utilizing cross-attention mechanisms where the PGV acts as a query, the network ensures that all branches learn within a unified semantic space from the outset, effectively bridging the gap between isolated perspectives.
MAPVLM Metric Learning
To overcome the limitations of Euclidean distance, the authors introduced Manifold Adaptation and Pairwise Variance-Locality Learning Metric (MAPVLM). This method learns a Symmetric Positive Definite (SPD) matrix to compute the Mahalanobis distance. The process involves:
- Noise Reduction: Utilizing Principal Component Analysis (PCA) to filter irrelevant features.
- Manifold Modeling: Employing heat-kernel-based affinity matrices to capture local neighborhood relationships.
- Feature Warping: The resulting SPD matrix weights feature dimensions by their discriminative power and accounts for inter-dimensional correlations, "warping" the feature space to better reflect the data's intrinsic manifold.
Experimental Results
The framework was rigorously evaluated on three major benchmarks: NCLT (long-term robustness), KITTI Odometry (urban loop closure), and Ford Campus (cross-environment generalization).
- Performance Benchmarks: On the NCLT dataset, the method demonstrated consistent improvements in Average Recall (AR@1, AR@5, AR@20) over baselines such as CVTNet. On KITTI, it achieved a state-of-the-art 91.9% AUC and 89.1% F1max.
- Generalization and Efficiency: The model exhibited strong zero-shot generalization on the Ford Campus dataset without fine-tuning. Furthermore, it maintains a real-time inference speed of approximately 30.54 ms, making it viable for deployment in autonomous systems.
- Metric Effectiveness: The Mahalanobis approach improved descriptor contrast by nearly 20% compared to Euclidean metrics, significantly enhancing the discriminability of similar locations.
Conclusion
This paper introduces a paradigm shift in LiDAR-based place recognition by moving from isolated multi-view processing to a holistic, pseudo-global fusion approach. By combining early contextual alignment with a learned Mahalanobis metric, the network effectively resolves the "information island" and "Euclidean trap" problems. Future work will focus on making the SPD matrix computation fully end-to-end and improving robustness in highly dynamic environments.
Summary updated