講演情報

[Data_Collection-03]Deep Learning-Based Photometric Stereo for Reliable Pavement Roughness Assessment

Benjamin Bringier1, *Majdi Khoudeir1, Minh-Tan Do2, Véronique Cerezo2, Damien Pilet3 (1. Xlim Institut (France), 2. Université Gustave Eiffel (France), 3. Ginger Group (France))

キーワード:

Pavement Roughness Measurement、Photometric Stereo、BRDF Modeling、Deep Learning

Road surface roughness plays a critical role in ensuring vehicle safety by maintaining adequate skid resistance. Traditional measurement methods, such as the Mean Texture Depth (MTD), remain widely used but suffer from significant drawbacks: they are time-consuming, require manual intervention, and lack repeatability, making them unsuitable for large-scale or rapid assessments. To address these limitations, this work introduces a portable optical system based on the photometric stereo technique combined with a multilayer perceptron (MLP) neural network. The proposed approach reconstructs the 2.5D geometry of pavement surfaces from six HDR images captured under varying illumination, enabling accurate estimation of surface normals and roughness parameters. The neural network is trained on a synthetic dataset generated through physically based rendering, incorporating diverse Bidirectional Reflectance Distribution Function (BRDF) models and optical phenomena such as shadows and inter-reflections. This design ensures robustness across different pavement types, including highly specular surfaces such as newly constructed road coatings, which are particularly challenging for traditional optical methods. Our acquisition device is lightweight and designed for field use, delivering complete measurements, including image capture, surface reconstruction, and ISO 25178 roughness parameter computation, in under one minute. A measurement campaign was conducted on road surfaces with varying macro- and microtexture characteristics. Results were compared against MTD values, demonstrating strong correlation, particularly for the volumetric Vvc criterion when excluding highly porous surfaces. While discrepancies occur for surfaces with hidden cavities, these are inherent to optical methods. Overall, the proposed system offers fast, reliable, and reproducible measurements, outperforming traditional techniques in efficiency and portability. This study highlights the potential of combining photometric stereo with deep learning for real-time pavement monitoring, paving the way for embedded solutions that enhance road safety and maintenance strategies.