Presentation Information

[Data_Collection-22]Application of Synthetically Trained Deep Learning Models in Chronological Pavement Surface Monitoring

*KRISHNA MODAK1, DI SU1, KAI XUE2, NAOTO YOKOZAWA3, TOMONORI NAGAYAMA1 (1. The University of Tokyo (Japan), 2. Smartcity Research Institute (Japan), 3. Public Works Research Institute (Japan))

Keywords:

Synthetic Dataset,Pavement Monitoring,Accelerated Pavement Testing,Chronological Analysis,Pavement Surface Deterioration

While traditional infrastructure maintenance has long relied on subjective and labor-intensive manual inspections, the industry is increasingly adopting automated computer vision systems to ensure objective, large-scale monitoring. However, the efficacy of these deep learning models depends heavily on high-quality training data, and acquiring pixel-perfect ground truth from the real world remains a significant bottleneck. Additionally, the pavement surface is generally subject to varying lighting conditions and surface degradation over time, which adds complexity to chronological analysis. This study introduces a comprehensive framework bridging virtual simulation and physical reality, using procedural generation to train deep learning models for complex damage assessment. The methodology begins by statistically analyzing real-world road imagery to extract color profiles for asphalt and distress patterns. These insights drive the procedural generation of pavement surface textures (e.g., alligator cracking, surface wear, weathered markings) in a dedicated texture authoring tool. Assets are then integrated into an automated game engine pipeline, which renders a diverse synthetic dataset with varied lighting conditions and perfect segmentation masks. This synthetic data serves as the training foundation for advanced deep learning models, including modified LoFTR networks for feature matching, enabling zero-shot transfer without manual real-world labeling. Though the resulting images do not achieve absolute photorealism, they offer a distinct advantage: controllable variability in lighting, crack patterns, and damage intensity. This flexibility allows for the simulation of rare edge cases that are financially or physically impractical to collect empirically, ultimately yielding models with superior generalization capabilities compared to those trained solely on limited real- world datasets. The application of this approach was implemented via an accelerated pavement cracking experiment on a controlled circular test track, where weighted trucks simulated extensive traffic fatigue. High-resolution video captured the chronological evolution of surface damage. Analysis using specific distress metrics like crack crosspoints identified a distinct non-linear deterioration pattern: pavement conditions stabilized initially but showed a sharp, critical failure onset after 33,000 truck passes. This application demonstrates the system’s ability to precisely quantify structural degradation and effectively identify the transition from early-stage wear to severe fracturing, solely relying on a model trained on a synthetically generated pavement surface image dataset.