A Hybrid Deep Learning Model for Concurrent Lane Detection and Pavement Distress Analysis

Abstract

The increasing volume of vehicular traffic and the progressive deterioration of road infrastructure demand intelligent and automated monitoring systems to enhance road safety and support timely maintenance. This paper presents AutoMed, a unified deep learning framework designed for real-time lane boundary detection and road surface defect identification from images and video streams. The proposed framework integrates Compact LaneNet for accurate lane segmentation with YOLOv8s for detecting road defects, including potholes and speed breakers, within a single processing pipeline. To improve robustness under challenging road conditions, a Hough Transform-based fallback mechanism is incorporated to maintain reliable lane detection when primary segmentation performance degrades. The models are trained using a synthetically augmented dataset with extensive optimization to improve generalization across diverse road geometries, weather conditions, and lighting environments. Furthermore, the framework is optimized for real-time deployment and supports multiple inference platforms through TensorFlow,ONNX, TorchScript, and OpenVINO, enabling efficient execution on resource-constrained devices. Experimental evaluation demonstrates that AutoMed achieves accurate, scalable, and computationally efficient road scene understanding, making it a promising solution for intelligent transportation systems, autonomous driving, and smart roadway infrastructure monitoring.