Abstract
Precise and long-term stable localization is essential in parking lots for tasks like autonomous driving or autonomous valet parking. Existing methods rely on fixed and memory-inefficient maps, which lack robust data association approaches and are unsuitable for precise localization or long-term maintenance. In this paper, we propose a novel mapping, localization, and map update system based on ground semantic features, utilizing low-cost cameras. We present a precise and lightweight parameterization method to establish improved data association and achieve accurate localization at the centimeter-level. Furthermore, we propose a novel map update approach by implementing high-quality data association for parameterized semantic features, allowing continuous map update and refinement during re-localization. We validate the performance of the proposed method in real-world experiments, achieving an average accuracy improvement of 5cm during registration. The generated maps consume only 450 KB/km and remain adaptable to evolving environments.
Introduction
Navigating parking lots presents significant challenges for autonomous vehicles. GPS signals are frequently unavailable in indoor or multi-story structures, and high-end LiDAR sensors remain cost-prohibitive for mass-market applications. While visual SLAM offers a low-cost alternative, it often struggles in "feature-poor" environments or under varying lighting conditions. Traditional semantic mapping methods typically produce static, memory-intensive maps that are difficult to update as the environment changes (e.g., repainting or layout modifications). LESS-Map addresses these gaps by providing an accurate, efficient, and adaptive mapping solution.
Methodology
The LESS-Map pipeline integrates mapping, localization, and map updates into a unified framework.
Semantic Parameterization
The system utilizes four surround fisheye cameras to detect ground features such as lanes, arrows, and stop lines. These features are projected into 3D space via Inverse Perspective Mapping (IPM). To achieve high efficiency, LESS-Map extracts contours from these markings and parameterizes them into 6-DOF line segments.
A key innovation is the use of contour normal vectors—vectors pointing from the interior of a marking to its exterior. This approach facilitates robust data association and enables "Normal-Assisted Odometry," which matches new observations to the map by ensuring consistent normal orientations. This method is significantly faster and more accurate than standard point-to-line matching algorithms like ICP.
Evolving Map Update
Unlike static maps, LESS-Map employs a Bayesian filter to continuously refine the map. As the vehicle re-localizes, the system fuses new observations to update the positions of existing features or add new ones in previously unmapped areas based on an "overlap ratio."
Experimental Results
The system was evaluated in real-world indoor and outdoor parking environments using a vehicle equipped with fisheye cameras, an IMU, and wheel encoders.
- Storage Efficiency: LESS-Map produces extremely compact maps, requiring only 450 KB/km. This is a significant improvement over ORB-SLAM2 (~55 MB/km) and AVP-SLAM (~2.8 MB/km).
- Localization Accuracy: The system achieved an average trajectory error of approximately 26cm, outperforming AVP-SLAM (35cm) and ORB-SLAM2 (57cm). The use of normal vectors contributed to a 5cm accuracy improvement during registration.
- Computational Performance: The matching process takes only 12.25 ms, nearly five times faster than conventional ICP-based methods (60 ms), making it ideal for real-time edge processing.
- Robustness: The system demonstrated consistent performance across various lighting conditions, including sunny, cloudy, and dusk scenarios.
Conclusion
LESS-Map provides a lightweight and adaptive solution for long-term localization in parking lots. By parameterizing ground markings and implementing a probabilistic update mechanism, the system maintains high accuracy while minimizing storage requirements. Future work will focus on extending the system to multi-agent environments, allowing a fleet of vehicles to collaboratively build and maintain a global semantic map.
Summary updated