A Low-Cost Vision-GPS Framework for the Unified Mapping of Vertical and Horizontal Road Assets Using Deep Learning

Area of Science:

• Computer Vision
• Geospatial Information Science
• Transportation Engineering

Background:

• Automated mapping of road assets is crucial for road safety and Intelligent Transportation Systems (ITS).
• Existing methods are often labor-intensive, lack a unified approach, or are proprietary, hindering accessibility and reproducibility.
• There is a need for an accessible, comprehensive framework for automated road asset identification.

Purpose of the Study:

• To present a comprehensive, low-cost framework for automated mapping of vertical traffic signs and horizontal road markings.
• To improve data efficiency and spatial diversity through a distance-aware frame extraction strategy.
• To address class imbalance for robust detection of all road asset categories.

Main Methods:

• Utilized a low-cost, vehicle-mounted action camera for data acquisition.
• Implemented a distance-aware frame extraction strategy to optimize data redundancy and spatial diversity.
• Employed deep learning models with segmentation-based annotation for accurate identification of vertical and horizontal road assets, including irregular markings.

Main Results:

• Achieved high performance in automated asset identification.
• Leading YOLO-based architectures attained an F1-score of 0.92 for vertical signage and 0.96 for horizontal markings.
• Successfully transformed raw visual data into structured, georeferenced information.

Conclusions:

• The developed framework provides a robust and accessible solution for automated road asset mapping.
• Facilitates the generation of High-Definition (HD) maps and digital inventories for proactive road maintenance and safety assessments.
• Supports road authorities in enhancing regional road safety and ITS development.