A Comprehensive Single-Camera Deep Learning Framework for Detecting Cracks and Potholes, Classifying Severity, and Quantifying Pothole Dimensions

Road distresses, such as potholes and multiple forms of cracking, significantly affect road safety, vehicle performance, and infrastructure quality. This study proposes a cost-effective and scalable framework leveraging monocular cameras and advanced computer vision models for detecting, classifying, and quantifying both cracks and potholes. Monocular cameras, widely available in vehicles as dashcams, provide an affordable alternative to expensive sensors, making this approach accessible in resource-constrained settings. The framework integrates YOLOv7-Tiny and YOLOv8-Nano models for accurate detection of road distresses, followed by severity classification using a ResNet50-based transfer learning model adhering to Indian Roads Congress (IRC) guidelines. For quantifying damage, the Segment Anything Model (SAM) is utilized for precise area estimation, and the Dense Prediction Transformer (DPT) enables depth estimation to calculate pothole volumes. Crack identification and pothole detection are jointly handled within the detection stage, ensuring consistent processing across different distress types. Comprehensive datasets, including manually labeled ground truth measurements, validate proposed system. Results show high detection accuracy, robust severity classification, and strong correlations between automated and ground truth areas and volume estimations, with coefficients of determination of 0.95 and 0.77, respectively. By enabling roadway agencies and municipalities to inspect and prioritize maintenance using low-cost tools, this framework supports scalable, data-driven pavement management even in resource-constrained environments. This study highlights potential of monocular cameras and deep learning for affordable, real-time road distress management, addressing critical gaps in traditional approaches reliant on costly sensors. Future work will focus on enhancing depth estimation accuracy for smaller damages and integrating real-time processing for deployment in resource-constrained environments.