Image-Based, Photometric-Stereo and Deep Learning Solutions for Road Surface Characterization

This thesis has its foundation in a cotutelle agreement involving the universities of Florence/Pisa in Italy and the University of Nantes in France. As a result, the doctoral candidate spent half of his experience on each institution, each contributing to its academic journey. The project was conceived as a response to develop a complementary system in addition to the current methods used for road surface characterization, with a particular focus around road surface texture and noise. The thesis is structured around three main topics: the detection of cracks using an image-based algorithm, photometric stereo techniques for three-dimensional reconstruction, and enveloped profiles for tire-road contact. The initial chapter centers on the identification of cracks by examining images sourced from publicly available datasets. Within this context, we present an enhancement to an existing algorithm, employing the principles of minimal path selection in conjunction with texture attributes extracted through the utilization of Gabor filters. A sequence of pre-processing steps are suggested, aimed at achieving shadow/lighting normalization and texture refinement, consequently leading to an improved detection of localized low points. By introducing a novel cost function that combines intensity levels with texture features, the resulting paths exhibit less dispersion. Additionally, we introduce a region-based segmentation technique that delineates two distinct areas. These areas are subsequently employed to establish threshold values, pivotal for the processes of elimination, skeletonization, and regional growth. The second chapter delves into photometric stereo methodologies for depth estimation, with an added innovation through the use of deep learning techniques. Initially, a close-range system was conceptualized and constructed, encompassing an exploration of parameters for calibration and optimal configuration. Subsequently, a solution was explored for semi-calibrated near point light sources, which was then tailored to our specific case study involving the replacement of single point light sources with a LED panel. Ultimately, we introduce a methodology aimed at enhancing outcomes by leveraging the capabilities of deep learning. The third part of this work focuses on the textural properties of the road surface, i.e., micro and macro-texture, and the procedure to obtain the envelopment profile for tire-road contact. Approaching the problem with a numerical solution, the vertical displacement of the elastic half-space at the frontier is computed. After this, a new methodology based on convolutional neural networks to estimate an initial guess of the contact areas between the two surfaces was developed. This initial estimation is then integrated into the original numerical methodology, resulting in reduced iterations and significantly accelerated convergence of the process. This thesis encompassed a diverse array of subjects, the exploration spanned across image-based algorithms, numerical optimizations, computer vision, contact mechanics, and delved into pragmatic concerns such as devising a lighting-box and calibrating the imaging system and lighting apparatus. While the interdisciplinary nature of this endeavor might initially appear disconcerting to readers, it's important to note that each facet examined in this study emerged as a response to a specific real-world problem in need of resolution. Furthermore, each topic underwent meticulous and comprehensive examination, resulting in well-founded propositions. The outcomes achieved for each individual topic represent noteworthy innovations within their respective fields, boasting practical implications that are readily discernible. Given the constraints of the relatively brief duration of the PhD program, a more extensive exploration of these topics couldn't be undertaken. However, it's worth highlighting that numerous possibilities for future development remain ripe for exploration beyond the scope of this program.