Abstract
Accurate three-dimensional road structures and models are of great significance to intelligent transportation applications, such as vehicle navigation, self-driving vehicles, and urban planning. This paper proposes an efficient and robust method to automatically extract structured road curbs from mobile laser scanning (MLS) data. The proposed method consists of three primary steps: efficient supervoxel generation, road curb detection, and driving free space estimation. By grouping ground points into supervoxels based on geometric characteristics, the method avoids the high computational cost of point-by-point processing while maintaining 3D precision. Experimental results on complex urban datasets from Shanghai, China, demonstrate that the method accurately and efficiently extracts road boundaries and driving free space.
Introduction
Creating high-definition (HD) maps for autonomous navigation requires precise 3D models of road structures. While satellite imagery and Aerial Laser Scanning (ALS) are common, they often suffer from low resolution and occlusions in urban canyons. Mobile Laser Scanning (MLS) provides significantly denser data but introduces challenges due to the massive, unorganized nature of the point clouds and uneven point densities.
Existing extraction methods typically fall into two categories: projection-based methods, which are computationally efficient but lose 3D precision, and point-based methods, which are accurate but computationally expensive. This paper introduces a "Goldilocks" solution: a supervoxel-based workflow that maintains 3D precision without the heavy computational overhead of traditional point-by-point analysis.
Methodology
The proposed pipeline follows a three-stage process to transform raw MLS data into vectorized road models.
1. Efficient Supervoxel Generation
Instead of calculating geometric features for every individual point, the system groups points into "supervoxels." The process begins with a 2D grid to identify candidate boundary voxels based on height differences. Points are then reallocated into supervoxels using a feature space difference formula that accounts for horizontal distance, projection density, and height. This approach preserves boundaries without requiring time-consuming surface normal calculations for every point.
2. Road Curb Detection
The system identifies supervoxels with significant height differences as potential curbs. To distinguish road curbs from other vertical structures like building bases, light poles, or tree trunks, the algorithm employs local height distribution histograms. This allows the system to identify the specific geometric signature of a curb (typically 10-20cm) versus taller vertical obstacles.
3. Driving Free Space Estimation
Once boundaries are established, the algorithm determines the "driving free space." It utilizes the vehicle's GNSS trajectory points as "seeds" for a region-growing algorithm. Adjacent horizontal supervoxels are merged if their height difference is minimal (less than 3cm), effectively mapping the flat, drivable road surface.
Experimental Results
The method was evaluated using two large-scale datasets collected in Shanghai, China, using an Alpha3D mobile laser scanning system:
- Xinchang Residential District: 10.9 km stretch containing approximately 500 million points.
- Zhangjiang Industrial District: 13.04 km stretch containing approximately 650 million points.
Key Performance Metrics
- Precision: The system successfully detected "tiny height step curbs" as small as 3cm.
- Efficiency: By using supervoxels as the processing unit, the method handled datasets exceeding 20GB while maintaining morphological integrity.
- Robustness: Qualitative analysis showed the algorithm successfully navigating complex intersections and curved roads where boundaries are not parallel to the vehicle's path.
Conclusion
This research provides a robust framework for automatic road structure detection. By leveraging trajectory data as a spatial constraint and supervoxels for efficient processing, the method overcomes common hurdles in MLS data analysis. Future work will focus on "gap-filling" logic to handle occlusions caused by parked vehicles and integrating the detection of horizontal road markings to create a complete vector model of the road environment.
Summary updated