Abstract
High-fidelity 3D mapping is essential for autonomous navigation, yet traditional dense point cloud representations impose significant burdens on storage, transmission, and multi-session consistency. This paper presents a lightweight mapping framework that replaces dense point clouds with semantic geometric primitives—specifically lines and planes. By optimizing these primitives on the Grassmannian manifold and employing a novel hybrid bundle adjustment (BA), the proposed method achieves a 1000x reduction in map storage while maintaining centimeter-level localization accuracy across multiple sessions.
Introduction
Standard LiDAR-based Simultaneous Localization and Mapping (SLAM) systems typically generate dense point cloud maps. While these maps provide rich environmental detail, they present three primary challenges for large-scale deployment:
- Storage and Bandwidth Constraints: Dense maps for a single session can reach gigabyte scales, complicating real-time sharing over vehicle-to-everything (V2X) networks.
- Multi-Session Inconsistency: Merging maps from different robot sessions or multiple robots often leads to geometric misalignments and high computational overhead.
- Redundancy: For localization purposes, many environmental details (e.g., vegetation) are unnecessary. Stable landmarks like poles and building facades provide sufficient geometric constraints.
To address these issues, this research proposes a "coarse-to-fine" centralized framework that prioritizes semantic geometric abstractions over raw point data.
Methodology
The framework utilizes a multi-stage approach to build and refine lightweight maps.
Semantic Primitive Extraction and Representation
The system extracts poles (represented as infinite lines) and building facades, roads, or fences (represented as infinite planes). To ensure optimization efficiency, the authors employ minimal parameterization: lines are represented with 4 parameters and planes with 3. This approach avoids the over-parameterization common in standard Euclidean forms, facilitating more stable convergence during optimization.
Grassmannian Place Recognition
A key challenge in multi-session mapping is recognizing previously visited locations without a prior pose estimate. The framework utilizes GraffMatch, a technique that treats groups of lines and planes as subspaces on a Grassmannian manifold. This allows for descriptor-free matching based purely on the geometric structure of the environment, enabling the system to merge disparate map sessions effectively.
Hybrid Bundle Adjustment (BA)
To ensure global consistency, the researchers developed a LiDAR-specific hybrid bundle adjustment. This process jointly optimizes the robot's trajectory and the parameters of the line and plane landmarks. By minimizing the distance between the raw LiDAR points and the corresponding semantic primitives across multiple sessions, the BA corrects odometry drift and ensures a cohesive global map.
Experimental Results
The framework was evaluated using the KITTI dataset (Sequences 00, 05, and 08) and the CARLA simulator (Town03).
Storage Efficiency
The most striking result is the reduction in map size. On KITTI Sequence 00, a standard point cloud map with 0.1m resolution requires approximately 1,773 MB. In contrast, the proposed semantic primitive map requires only 0.77 MB—a reduction of over 2,300 times.
Localization and Consistency
Despite the drastic reduction in data, localization accuracy remains high. Across the tested KITTI sequences, the system maintained a localization error between 2.3 cm and 6.0 cm. Furthermore, the hybrid BA significantly reduced the Relative Pose Error (RPE) compared to the A-LOAM baseline, demonstrating that optimizing geometric primitives effectively mitigates odometry drift.
Robustness
In CARLA simulations, the system demonstrated resilience to environmental noise. Even with 20% incorrect semantic labels and added Gaussian noise, the framework successfully merged ten different driving sessions into a consistent global map.
Discussion and Conclusion
By focusing on semantic classes like poles and buildings, the system naturally filters out dynamic noise from pedestrians and vehicles. The use of infinite primitives provides a higher-level abstraction that simplifies the mathematics of map merging and reduces storage requirements to a fraction of traditional methods.
However, the system's performance is currently tied to the availability of geometric primitives. In unstructured environments, such as forests or open fields, the lack of lines and planes may limit its effectiveness. Future work will focus on extending this framework to crowd-sourced mapping, enabling a fleet of low-cost vehicles to contribute lightweight map patches to a central server for real-time, city-scale map maintenance.
Summary updated