Architecture for pavement pothole evaluation using deep learning, machine vision, and fuzzy logic

Roads are critical for economic growth and trade but are constantly degraded by heavy traffic and adverse weather, leading to potholes that compromise safety. Traditional detection methods, like manual inspections, are labor-intensive, costly, and prone to errors. Existing automated systems also struggle with false positives, particularly in challenging conditions involving shadows, stains, or other environmental interferences. This research presents an architecture for detecting, measuring, and evaluating potholes, as well as generating maintenance recommendations. We integrate 2D images and 3D point clouds captured using the Intel RealSense D435i camera, generating a dataset of 583 images-299 containing potholes and 234 depicting environmental noise on various pavements. Each image is labeled through semantic segmentation and paired with corresponding point clouds. The architecture utilizes transfer learning with a Segformer network, achieving high detection performance with a Recall of 90.87 %, Precision of 90.01 %, Accuracy of 86.8 % F1 Score of 90.433 %, and a loss of 0.0431. The method achieves an IOU of 85.872 %, ensuring accurate diameter estimation, which contrasts with studies using lower IOU values where pothole dimensions are often underestimated due to incomplete detection. Our architecture provides reliable contour detection, facilitating the integration of image data and point clouds to estimate pothole dimensions with a depth estimation error of 5.94 mm. A fuzzy logic system processes these measurements to assess repair urgency and recommend appropriate repair techniques.