Abstract
Road extraction plays a significant role in the production of high-definition maps (HD maps). This paper presents a novel boundary-enhanced supervoxel segmentation method for extracting road edge contours from Mobile Laser Scanning (MLS) point clouds. The proposed method first leverages normal feature judgment to obtain 3D point clouds' global geometric information, then clusters points according to an existing method with global geometric information to enhance the boundaries. Finally, it utilizes the neighbor spatial distance metric to extract the contours and drop out existing outliers. The proposed method is tested on two datasets acquired by a RIEGL VMX-450 MLS system that contain major point cloud scenes with different types of road boundaries. The experimental results demonstrate that the proposed method provides a promising solution for extracting contours efficiently and completely. Results show that the precision values are 1.5 times higher than existing methods when the recall value is 0 for both tested road datasets.
Introduction
The creation of High-Definition (HD) maps is a cornerstone of autonomous driving technology. A critical component of these maps is the precise extraction of road boundaries from 3D point clouds. While Mobile Laser Scanning (MLS) provides dense geometric data, standard "supervoxel" methods—which group nearby points into small, meaningful volumes—often "smooth over" the very edges they are intended to detect.
Existing approaches frequently rely on additional data like color or intensity, which are not always available or reliable in diverse lighting and weather conditions. Without these attributes, supervoxel segmentation often fails to respect the subtle geometric shifts at road curbs, leading to incomplete or noisy edge extraction. This paper introduces a boundary-enhanced supervoxel method designed to precisely extract road edges by integrating global geometric constraints into the clustering process.
Methodology
The proposed method utilizes a three-stage pipeline that prioritizes "global spatial judgment" to maintain sharp road edges.
- Geometric Information Judgment: The system calculates normal vectors for every point using tangent planes and eigenvalues. By applying height and angle thresholds, it identifies points likely belonging to vertical structures, such as curbs.
- Enhanced Point Clustering: Building on established clustering techniques, the algorithm is modified to use the previously calculated geometric information. This ensures that points near road edges are merged into supervoxels that align with the boundary rather than crossing over it, preserving the edge's integrity.
- Boundary Contour Extraction: The method employs a $k$-Nearest Neighbor (kNN) search to discard large, irrelevant segments and a Radius Nearest Neighbor (RNN) search to filter out outliers. This leaves behind a clean, precise road contour based on the supervoxel labels.
Experimental Setup and Datasets
The researchers validated the method using a combination of public benchmarks and custom-captured data:
- IQmulus & TerraMobilita (IQTM): An urban point cloud dataset featuring 12 million manually labeled points from the streets of Paris.
- Semantic 3D: A large-scale outdoor classification benchmark used to test the method's versatility across different environments.
- Ring Island Road Datasets: Real-world data acquired specifically for this study using a RIEGL VMX-450 MLS system, which includes dual laser scanners, IMU, and GPS.
Results and Discussion
The proposed method was compared against industry standards, including VCCS and facet segmentation methods.
Key Insights
- Geometry over Color: The results demonstrate that RGB color or intensity data are not strictly necessary for edge detection if the underlying "global spatial structure" is leveraged effectively.
- Boundary-Preserving Supervoxels: Unlike traditional algorithms that aim for uniform shapes, this method produces supervoxels that are denser and more irregularly shaped near boundaries, capturing high-frequency changes at curbs.
- Outlier Resilience: The two-step extraction process (kNN followed by RNN) effectively ignores noise, such as stray points on the road surface or sidewalk.
Performance Metrics
On the tested MLS datasets, the precision was 1.5 times higher than the region-growing method at the start of the recall curve (recall = 0). In the IQTM dataset, the proposed method generated significantly more supervoxels (9,676) compared to VCCS (105) or Lin’s method (193), allowing it to trace complex boundaries with much higher fidelity. Qualitative analysis shows that while other methods produced jagged or incomplete lines, this method extracted smooth, continuous road edges.
Conclusion
The boundary-enhanced supervoxel method provides a robust foundation for HD map production by accurately extracting road edges from MLS data. While highly effective for environments with distinct curbs or height changes, the method's primary limitation is its reliance on vertical edges. Future work will focus on improving performance in environments with flat road-to-shoulder transitions and areas with heavy interference from dense vegetation or complex building structures.
Summary updated