Files
Abstract
The survey and monitoring of pavement distress are vital for pavement safety, maintenance, and evaluation. Accurate detection and quantification of pavement distress with complex physical topologies can deliver sufficient information for the assessment of pavement quality and the prevention of further pavement distress. In this study, two pavement image databases were created. The first database comprises over 19,237 images of seven primary pavement distress types, collected from digital cameras, open-source datasets, unmanned aerial vehicles, and infrared cameras. The second database is a processed open-source street view imagery dataset consisting of 25,540 images. The study evaluates multiple machine learning methods on multi-type pavement distress detection, and compares their performance in object detection and classification tasks. The study further explores attention-based improvements in You Only Look Once (YOLO) networks for pavement distress detection. The YOLO version 8 framework is optimized using multi-head self-attention (MHSA), and selective kernel (SK) attention mechanisms. Comparative analyses of different YOLO versions and optimized YOLO models are conducted to identify the most effective techniques. The dissertation then details the development of an integrated pavement distress detection system application which offers real-time distress detection, categorization, GPS positioning, and detection confidence. To enhance efficiency and robustness, the research also focuses on spatial and temporal anomaly analysis of pavement condition data to improve data integrity and prediction accuracy. Furthermore, the enhanced data serve as the input in a novel pavement roughness estimation framework. This framework uses street-view images and pavement International Roughness Index (IRI) data to explore the correlation between digital images and IRI. Based on a 40-meter per image criterion, 5,108 images are processed with the Segment Anything Model (SAM) to isolate pavement pixels and subsequently used to train models built on a ResNet backbone with Simple Weighted Feature Fusion (SWFF) module and an attention-augmented SWFF (SWFF-CBAM) module. The SWFF model attains a maximum R² of 0.972, while the SWFF-CBAM model reaches 0.944 in fewer epochs. Tests across various pavement types confirm the method's effectiveness in assessing pavement roughness on different pavement surfaces.