Novel Methodology to Recover Road Surface Height Maps from Illuminated Scene through Convolutional Neural Networks

Area of Science:

• Civil Engineering
• Computer Science
• Artificial Intelligence

Background:

• Road surface properties significantly influence pavement service life.
• Contactless monitoring techniques, like laser profilometers, are precise but costly and lack spatial information.
• There is a need for affordable, high-precision road surface analysis tools.

Purpose of the Study:

• To develop a fast, low-cost system for recovering road surface height information using commercial cameras.
• To leverage Convolutional Neural Networks (CNNs) for 3D road surface reconstruction.
• To demonstrate the feasibility of AI-based road surface monitoring.

Main Methods:

• A dataset was created using a custom-built photometric stereo rig with four light sources.
• Commercial camera images were captured for road surface samples with markers.
• Ground truth data was obtained using a laser profilometer for network training and validation.
• A CNN was trained and validated using the ad hoc dataset, evaluating three loss functions.

Main Results:

• The developed system successfully reconstructed road surface height information.
• Binary Cross Entropy loss function demonstrated the best performance for the reconstruction task.
• The AI-based methodology proved effective in overcoming the limitations of traditional methods.

Conclusions:

• A feasible, low-cost road surface monitoring system using commercial cameras and AI has been developed.
• The proposed CNN-based approach offers a cost-effective and efficient alternative to laser profilometry.
• This methodology paves the way for widespread adoption of AI in pavement management and maintenance.