Computer vision and convolutional neural networks applied in pavement evaluation

Abstract

The implementation of new technologies in pavement management systems (PMS) is crucial, particularly through computer vision and machine learning to develop automatic pavement evaluation systems. This research presents a deep convolutional neural network (CNN) model capable of classifying various pavement types, including concrete, asphalt, interlocking, cobblestone/stone, and unpaved roads, achieving nearly 100% accuracy. Additionally,the study explores image-based distress diagnosis using multi-label classification (MLC) models with ResNet-50 architecture, achieving a 96% accuracy rate and a 91% F1-score, effectively identifying pavement distress from network-level video surveys. This approach reduces the need for expensive sensors, offering a cost-effective alternative for comprehensive road condition assessments. Continuous and extensive examination of pavement surfaces is vital for PMS effectiveness, and automated assessments using CNNs offera feasible solution by quickly and accurately identifying defects, thereby reducing manual evaluation costs. The research proposes a Simplified Pavement State Index (SPS) using 2D images from low-cost cameras, demonstrating high similarity in prioritizing segments for intervention compared to the traditional Global Severity Index (IGG). Automating data collection through CNNs provides increased efficiency, enabling continuous monitoring and timely maintenance. Despite the initial investment and challenges in ensuring system robustness, the use of action cameras and smartphones presents a low-cost alternative supporting network-level PMS initiative. This study emphasizes the importance of collaboration between academia and highway management agencies for continuous improvement and underscores the potential of automated assessments to revolutionize pavement management, providing rapid, and comprehensive evaluations, ultimately aiding in effective maintenance and extending pavement service life.