Abstract
High Accuracy Driving Maps (HADMs) are the core component of Intelligent Drive Assistant Systems (IDAS), which can effectively reduce traffic accidents due to human error and provide more comfortable driving experiences. Vehicle-based mobile laser scanning (MLS) systems provide an efficient solution to rapidly capture three-dimensional (3D) point clouds of road environments with high flexibility and precision. This paper proposes a novel method to extract road features (e.g., road surfaces, road boundaries, road markings, buildings, guardrails, street lamps, traffic signs, roadside-trees, power lines, vehicles and so on) for HADMs in highway environments. Quantitative evaluations show that the proposed algorithm attains an average precision and recall of 90.6% and 91.2%, respectively. Results demonstrate the efficiency and feasibility of the proposed method for large-scale road feature extraction.
Introduction
For autonomous vehicles to navigate safely, they require High Accuracy Driving Maps (HADMs) that serve as a high-fidelity "digital twin" of the road environment. These maps must include precise locations for lane markings, guardrails, traffic signs, and even overhead power lines. While Mobile Laser Scanning (MLS) systems can capture the necessary 3D data, the resulting point clouds are massive, unorganized, and noisy. Traditionally, feature extraction has been a bottleneck, requiring either labor-intensive manual processing or struggling with "intensity range effects"—where the brightness of a laser return fluctuates based on distance, complicating the identification of painted road markings. This research presents an automated pipeline to bridge this gap.
Methodology
Automated Extraction Pipeline
The proposed framework follows a multi-stage process starting from raw MLS data to final object recognition. The pipeline involves ground filtering, road surface extraction, and a parallel track for roadside object segmentation and classification.
Road Surface and Marking Extraction
To overcome the limitations of simple intensity thresholds, the authors developed an intensity gradient-based algorithm.
- Surface Detection: RANSAC is used to fit planes to ground points, identifying the road surface relative to the vehicle's trajectory.
- Marking Extraction: Instead of absolute intensity, the system calculates the intensity gradient for each point. By applying K-means clustering to these gradients, the algorithm identifies the sharp "edges" of road markings. This approach makes the detection immune to brightness fluctuations caused by distance or sensor angle.
Roadside Object Extraction
For discrete objects like trees and poles, a "segment-then-classify" approach is employed:
- Segmentation: Points are classified by local geometry (linear, planar, or spherical) and grouped using a multi-rule region growing algorithm.
- Multi-Level Feature Classification: A Support Vector Machine (SVM) classifies objects based on four data levels: point-based local shapes (Fast Point Feature Histograms), segment-based size/orientation, object-based global shapes (Viewpoint Feature Histograms), and context-based relationships (e.g., height consistency or regular spacing of street lamps).
Results and Performance
The method was tested on a 6.0 km highway dataset containing 167.8 million points captured by a VMX-450 MLS system. The system achieved an average precision of 90.6% and an average recall of 91.2%.
Buildings and guardrails were identified with the highest precision (>93%), while trees and cars were slightly lower (83-86%) due to their irregular shapes and potential occlusions. The entire dataset was processed in just 12.6 minutes on a standard high-end laptop, demonstrating significant efficiency for large-scale mapping.
Key Insights
- Contextual Importance: Geometry alone often fails to distinguish similar objects (e.g., utility poles vs. thin trees). Incorporating context-based features like "distance to road boundary" significantly improves classification.
- Gradient Robustness: Focusing on intensity gradients rather than absolute values solves the "dimming" problem inherent in LiDAR data for distant objects.
Conclusion
This automated pipeline provides a robust solution for generating HADMs from dense MLS data. While the system performs exceptionally well, future work will address segmentation errors—specifically over- and under-segmentation—and explore the fusion of spectral (camera) data to further refine the recognition of complex roadside objects.
Summary updated