Abstract
Point cloud based place recognition plays an important role in mobile robotics. In this paper, we propose a weighted aggregation method from structure information adaptively for point cloud place recognition. Firstly, to preserve the prior distributions and local geometric structures, we fuse learned hidden features with handcrafted features in the beginning. Secondly, we further extract and aggregate adaptively weighted features concerning density and relative spatial information from these fused features, named Weighted Aggregation with Density Estimation (WADE) module. Then, we conduct the WADE block iteratively to group the latent manifold structures. Finally, comparison results on two public datasets Oxford Robotcar and KITTI show that the proposed approach exceeds the comparison approaches on recall rate averagely 7% - 8%.
Introduction
Place recognition—the ability of a robotic system to identify a previously visited location—is a fundamental component of autonomous navigation and Simultaneous Localization and Mapping (SLAM). While vision-based systems are widely used, they are often susceptible to environmental changes such as lighting fluctuations, seasonal variations, and weather conditions. 3D LiDAR point clouds offer a more robust alternative, providing precise geometric information regardless of ambient light.
However, LiDAR data presents unique challenges: point clouds are inherently unordered, sparse, and exhibit significant density variations depending on the distance from the sensor and the nature of the objects scanned. Current deep learning approaches often struggle to account for these non-uniform distributions or overlook valuable "prior" geometric knowledge.
Methodology: The WADE-Net Framework
WADE-Net introduces a novel architecture designed to bridge the gap between traditional geometric processing and modern neural networks. The framework operates in three primary stages:
1. Hybrid Feature Fusion
Rather than relying solely on raw XYZ coordinates, WADE-Net begins by fusing learned features with handcrafted geometric priors. These priors include:
- Range Value: The Euclidean distance from the sensor.
- Density Value: A measure of local point concentration.
- Normal Vectors: Indicators of surface orientation. By combining these with features learned via a Multi-Layer Perceptron (MLP), the network starts with a richer representation of the local environment.
2. The WADE Module
The core innovation is the Weighted Aggregation with Density Estimation (WADE) module. It processes fused features through two specialized branches:
- D-Branch (Density Estimation): This branch utilizes a Gaussian kernel to estimate local point density. It generates a "density factor" that allows the network to prioritize structurally significant areas (like corners or complex objects) over redundant ones (like flat walls).
- W-Branch (Weighting): This branch learns relative spatial relationships by transforming neighboring points into a local coordinate system, ensuring rotation-invariant feature extraction.
3. Iterative Manifold Grouping
The WADE module is applied iteratively (typically three times). At each step, the point cloud is downsampled using Farthest Point Sampling (FPS). This iterative process allows the network to capture latent manifold structures at multiple scales, moving from fine-grained local details to broader environmental contexts.
Finally, the local features are aggregated into a compact global descriptor using a NetVLAD layer. The model is trained using a Lazy Triplet Loss, which optimizes the descriptor space to ensure that similar locations are clustered together while distinct locations are pushed apart.
Experimental Results
The authors evaluated WADE-Net on the Oxford RobotCar and KITTI datasets, both of which are benchmarks for large-scale outdoor place recognition.
Performance Metrics
WADE-Net demonstrated state-of-the-art performance across multiple metrics:
- Recall Rates: On the Oxford dataset, WADE-Net achieved an 82.22% Top-1 recall, outperforming LPD-Net (76.62%) and PointNetVLAD. In several KITTI sequences, the model achieved near-perfect recognition with recall rates exceeding 99%.
- Efficiency: Despite its sophisticated weighting mechanism, WADE-Net is highly efficient. It contains only 2.009 million parameters, which is approximately one-fifth of the size of LPD-Net (~10M parameters). It also requires significantly fewer GFLOPs (2.453 vs. 6.859).
Discussion and Conclusion
The primary strength of WADE-Net lies in its "Hybrid Intelligence"—the strategic combination of handcrafted geometric features with adaptive deep learning. By explicitly modeling density, the network can effectively navigate the sparsity of outdoor LiDAR data.
While the model is computationally lean, the current implementation of handcrafted feature extraction (normals and density) is performed on the CPU, leading to an inference latency of 23.10ms. Future iterations could migrate these calculations to the GPU to achieve true real-time performance. Overall, WADE-Net represents a significant step forward in robust, efficient 3D place recognition for mobile robotics.
Summary updated