AI assisted pothole detection and depth estimation
Eshta Ranyal, Ayan Sadhu, Kamal Jain
Abstract
AI-assisted engineering solutions integrated with commercial RGB sensors and computationally intensive Graphical Processing Units (GPUs) promise a low-cost solution, to prevent deterioration of premature pavement disintegration. Potholes a common pavement distress are a severe threat to road safety and demand time and cost-effective state-of-the-art technologies for road inspection and condition monitoring. An intelligent pavement pothole detection system is proposed in this study by modifying the single stage CNN architecture-RetinaNet to detect potholes and perform metrological studies using 3D vision. The photogrammetric technique of structure from motion based on image frames extracted from pavement video recordings is used to model the 3D point cloud structure of potholes to assess the severity of the detected potholes as a function of its depth and is integrated with the CNN based pothole detection system. High F1 scores on benchmark dataset with a high value of 0.98, validate the model’s performance. A mean error below 5% is obtained on the measured depths thus promising an intelligent and practical solution to be implemented as part of a potential pavement health assessment system for future practice.