Abstract
This study aims at building a robust semi-automated pavement marking extraction workflow based on the use of mobile LiDAR point clouds. The proposed workflow consists of three components: preprocessing, extraction, and classification. In preprocessing, the mobile LiDAR point clouds are converted into the radiometrically corrected intensity imagery of the road surface. Then the pavement markings are automatically extracted with the intensity using a set of algorithms, including Otsu’s thresholding, neighbor-counting filtering, and region growing. Finally, the extracted pavement markings are classified with the geometric parameters using a manually defined decision tree. Case studies are conducted using the mobile LiDAR dataset acquired in Xiamen (Fujian, China) with different road environments by the RIEGL VMX-450 system. The results demonstrated that the proposed workflow and our software tool can achieve 93% in completeness, 95% in correctness, and 94% in F-score when using Xiamen dataset.
Introduction
For autonomous vehicles (AVs) to navigate safely, they require "prior maps"—highly detailed 3D semantic maps that provide precise locations of lanes, stop lines, and traffic symbols. While Mobile Laser Scanning (MLS) is the preferred method for collecting this data, the resulting point clouds are often massive and noisy. A significant challenge lies in the LiDAR "intensity" (brightness) returns, which vary based on the distance and angle of the scanner, often leading to data gaps or misidentification of markings. This paper proposes a robust, semi-automated workflow to bridge the gap between raw, messy point clouds and clean, categorized semantic maps.
Methodology
1. Preprocessing & Intensity Correction
The first stage involves separating the road surface from "non-ground" objects like trees and poles using a voxel-based upward growing method. To address uneven brightness, the researchers apply a linear regression model based on the cosine of the scan angle. This normalization "flattens" the intensity data, ensuring that pavement markings stand out uniformly against the asphalt regardless of the scanner's position.
Following normalization, the 3D points are converted into a 2D digital terrain model (DTM) using Inverse Distance Weighted (IDW) interpolation.
2. Road Marking Extraction
With the intensity normalized, the system employs Otsu’s thresholding to separate paint from the road surface. To refine the results, neighbor-counting filters and region-growing algorithms are used to eliminate small clusters of noise and ensure the continuity of extracted features.
3. Classification via Decision Tree
The extracted shapes are classified using a hierarchical decision tree based on geometric parameters such as area, perimeter, orientation, and width. For instance, the system distinguishes longitudinal lane lines from transverse stop lines based on orientation variance. It further categorizes segments by width; segments wider than 45 cm are flagged as "wide" markings, such as arrows or zebra crossings.
Experimental Evaluation
The study utilized a dataset from the Island Ring Road in Xiamen, China, collected via a RIEGL VMX-450 system. The survey covered 10 km with an average point density of 7,000 points/m². Seven distinct road sections, including complex intersections and curved roads, were used for evaluation.
The workflow achieved high performance across all metrics:
- Completeness: 93%
- Correctness: 95%
- F-Score: 94%
Comparative tests indicated that this method outperforms several established techniques and remains competitive with state-of-the-art deep learning methods while maintaining higher processing efficiency through dimensionality reduction (converting 3D clouds to 2D intensity images).
Discussion and Conclusion
The core strength of this approach is the intensity normalization, which allows simple global thresholding to be remarkably effective. By leveraging standardized geometric parameters (national road standards), the system can precisely classify complex symbols without the need for massive training datasets.
However, limitations remain. The system struggles with heavily decayed or faded markings where intensity returns are low. Additionally, low point density in certain areas can lead to extraction failures. Future work will focus on incorporating mathematical morphology to better handle complex characters and using spatial contextual logic to rectify fragmented lane lines.
Summary updated