PotholeNet: An Integrated Deep Learning Framework for Real-Time Road Pothole Detection and Spatial Dimension Estimation

Abstract

Road infrastructure deterioration, particularly pothole formation, is a persistent hazard to transportation safety and vehicle maintenance. Existing manual inspection methods are labour-intensive and inconsistent, while accelerometer-based approaches cannot quantify the physical dimensions needed to prioritise repair. This paper proposes PotholeNet, an integrated framework whose target design combines a customised YOLOv8 detection branch with a MiDaS-derived monocular depth branch to deliver simultaneous pothole localisation and metric-scale dimension estimation from a monocular camera feed. Because no ethically-accessible, camera-calibrated pothole dataset with ground-truth physical dimensions was available in this study's software-only development environment, and because training a full YOLOv8/MiDaS stack requires GPU/CUDA resources beyond that environment, the paper implements and genuinely evaluates a lightweight, fully-reproducible pilot instead of assuming the full architecture's performance: a calibrated pinhole/IPM camera model was used to generate a 370-image synthetic road dataset (230 pothole images across four lighting conditions, plus oil-stain, shadow and clean-road negatives) with exact ground-truth position, size and depth; a two-stage classical-CV-plus-trained-classifier pipeline performs detection; and length/width are recovered via geometric back-projection while depth is recovered via a contrast-to-depth regression. On a held-out test split, the pipeline achieves image-level precision of 75.0%, recall of 68.6%, and F1 of 71.6%, with an ablation study confirming that the trained classifier stage improves F1 by 5.7 points and accuracy by 16.1 points over classical thresholding alone. Depth estimation achieves a mean absolute error of 3.58 cm; length and width estimation show a systematic underestimation bias (MAE 16.8 cm and 18.0 cm respectively) traced to core-only segmentation of the pothole region, an honestly-reported limitation of the classical detection stage rather than the geometric back-projection itself. These results are presented as a transparent proof-of-concept for the detection and geometric-estimation mechanics of the proposed architecture, with a concrete roadmap toward the full YOLOv8+MiDaS implementation on real, camera-calibrated field data.