TL;DR
This paper introduces NDT-Map-Code (NDT-MC), a novel 3D global descriptor for lidar SLAM that excels at loop closure detection, particularly in challenging underground parking environments. By leveraging the efficient Normal Distribution Transform (NDT) representation and encoding geometric shape, height, and entropy, NDT-MC achieves state-of-the-art performance with reduced computational and storage costs, making it suitable for real-time applications and crowd-sourced mapping.
What Problem Does This Paper Address?
Loop closure detection is a critical component of Simultaneous Localization and Mapping (SLAM) systems, enabling them to identify previously visited locations and correct accumulated drift. Existing methods for lidar-based loop closure often rely on dense point clouds and 360-degree lidar scans, which are computationally expensive and require significant storage and transmission bandwidth. Furthermore, many current approaches are optimized for outdoor, on-road scenarios and struggle with the unique challenges of indoor environments like underground parking lots, such as repetitive structures, dynamic objects, and limited visibility. This paper aims to bridge this gap by proposing a lightweight, efficient, and robust global descriptor suitable for both indoor and outdoor environments, with a particular focus on the demanding conditions of underground parking.
Methods
The core of the proposed method is the NDT-Map-Code (NDT-MC) descriptor. The process begins by converting raw lidar point clouds into an NDT representation. This involves dividing the 3D space into uniform grids and fitting a multivariate Gaussian to the points within each cell, resulting in a mean and covariance matrix for each NDT cell.
NDT-MC then extracts features from these NDT cells in a multi-step process:
- Geometric Shape Classification: Each NDT cell's shape is classified into categories like plane, ellipsoid, sphere, and line. This is achieved by analyzing the eigenvalues of the covariance matrix and using a derived index
g. - Entropy Calculation: The entropy of the points within each NDT cell is calculated, providing a measure of the distribution's randomness.
- Polar-Range-Height (PRH) ROI Partitioning: The 3D space is partitioned into regions of interest (ROIs) using polar coordinates (range and angle) and height layers. This is a key departure from some methods that might only use bird's-eye views.
- Height-Layer Encoding: The scene is segmented into multiple layers based on height, and the geometric shape and entropy of NDT cells within each layer are encoded. This is particularly important for underground parking where ceiling height is consistent, making height-based discrimination crucial.
- Descriptor Formulation: The final NDT-MC descriptor is constructed by combining the encoded geometric and entropy features from each ROI, weighted by height.
For matching, the paper proposes a two-stage approach:
- Geometric Key (GK): A histogram-based "geometric key" is generated from the shape classification index
gto enable fast retrieval using a kd-tree. - Sector Key and Similarity Score: If a candidate keyframe passes the initial GK check, a "sector key" is computed by averaging descriptor columns. The similarity between descriptors is then calculated using the correlation coefficient with column-wise shifting to account for rotation.
Datasets
The paper evaluates its method on two distinct datasets:
- NIO Underground Parking-Lot Dataset: This is a newly collected dataset specifically designed for evaluating loop closure in underground parking scenarios. It consists of eight testing datasets from four different parking lots, captured using an ET7 model equipped with an Innovusion Falcon LiDAR. Ground truth trajectories are generated using a NovAtel IMU-ISA-100C.
- KITTI Dataset: The widely-used KITTI dataset is employed for evaluating loop closure detection in on-road driving scenarios. The Velodyne HDL-64E lidar sensor data is used.
Key Insights and Ideas
- NDT as a Scalable Foundation: The use of NDT representation is crucial for scalability and efficiency. It significantly reduces the point cloud size (e.g., 0.36% of raw data for 2m resolution), making it ideal for crowd-sourced mapping and real-time processing.
- Height-Layer Encoding for Indoor Environments: Explicitly encoding height information in layers is a critical innovation for indoor environments like parking lots, where ceiling height is consistent and can be a strong discriminative feature, unlike methods that might rely solely on maximum height.
- Geometric Shape and Entropy Synergy: Combining explicit geometric shape classification of NDT cells with their entropy provides a richer and more robust descriptor than either feature alone.
- Adaptability to Front-View Lidars: The method is designed to work with front-view lidars, addressing a limitation of many existing approaches that assume 360-degree coverage.
- Real-time Integration: The successful integration of NDT-MC into a full SLAM system (NDTMC-LIO-SAM) demonstrates its practical applicability in real-time scenarios.
Key Results
- Superior Performance on NIO Dataset: NDT-MC consistently achieves the highest F1 Score and Extended Precision (EP) across all eight testing trajectories on the NIO underground parking-lot dataset, outperforming methods like Scan Context, Lidar Iris, NDD, and STD.
- Competitive Performance on KITTI: On the KITTI dataset, NDT-MC achieves the highest F1 Score in several sequences (00, 02, 05, 06) and the highest EP in others (05, 06, 07, 08), demonstrating strong generalizability.
- Runtime Efficiency: Table IV shows that NDT-MC has significantly faster descriptor extraction and query times compared to Scan Context and Lidar Iris on the KITTI dataset, making it suitable for real-time applications.
- Integrated SLAM Performance: The NDTMC-LIO-SAM system shows comparable or superior trajectory accuracy to SC-LIO-SAM, particularly on challenging sequences like KITTI 08.
Limitations and Future Work
- Limitations:
- The paper notes that some methods (like STD) did not achieve 100% precision at various thresholds, preventing the derivation of the full RP100 value for Extended Precision calculation.
- While NDT-MC performs well, the paper acknowledges that "dense point cloud is not applicable" for some comparisons on the NIO dataset, implying that performance might differ if dense point clouds were available and utilized by all methods.
- Future Work:
- The paper suggests integrating the proposed global descriptor into a full SLAM system, which they have done with NDTMC-LIO-SAM, and making the code publicly available.
- The generalizability of the descriptor across a wider variety of indoor and outdoor environments could be further explored.
- Investigating the impact of different NDT cell resolutions and parameter tuning for various environments could be a future direction.
Summary updated