Reliable and timely perception of road surface conditions is a fundamental requirement in intelligent transportation systems (ITS), as it directly affects traffic safety, infrastructure maintenance, and the operation of connected and autonomous vehicles. Vision-based pothole detection has emerged as a practical solution due to its low sensing cost and deployment flexibility; however, existing deep learning approaches often struggle to achieve a satisfactory balance between detection accuracy, robustness to scale variations, and real-time performance on resource-constrained platforms. This study presents Partial Group-You Only Look Once (PG-YOLO), a lightweight real-time vision framework designed for road infrastructure monitoring in ITS. Built upon a compact one-stage detector, the proposed framework introduces a Partial Multi-Scale Feature Aggregation (PMFA) module to enhance the representation of small and irregular potholes under complex road conditions, as well as a Grouped Semantic Enhancement Attention (GSEA) module to improve high-level semantic discrimination with limited computational overhead. The framework is specifically designed to meet the low-latency and low-complexity requirements of vehicle-mounted and roadside sensing devices. Experimental evaluations conducted on a mixed road damage dataset demonstrate that the proposed approach achieves consistent improvements in detection accuracy while reducing model parameters and maintaining real-time inference speed. Compared with the baseline model, PG-YOLO improves precision, recall, and detection stability under challenging illumination and scale variations, while remaining suitable for edge deployment. These results indicate that the proposed framework can serve as an effective perception component for ITS, supporting continuous road condition awareness and data-driven maintenance and safety management.
Open Access
News 
Journal Metrics
Journal Overview
Associated Journals