Durham School of Architectural Engineering and Construction
Department of Construction Engineering and Management: Dissertations, Theses, and Student Research
First Advisor
Chun-Hsing Ho
Date of this Version
Summer 7-2025
Document Type
Thesis
Citation
A THESIS Presented to the Faculty of The Graduate College at the University of Nebraska In Partial Fulfillment of Requirements For the Degree of Master of Science
Major: Construction Engineering and Management
Under the Supervision of Professor Chun-Hsing Ho
Lincoln, Nebraska
July 2025
Abstract
Adviser: Chun-Hsing Ho
The long-term, sustainable assessment of bicycle road conditions is essential for ensuring cyclist safety and promoting equitable infrastructure management. While traditional detection methods based on manual thresholds have demonstrated some effectiveness, they carry limitations due to the diversity of pavement materials, cyclist behaviors, and sensor platforms. Although recent advances in deep learning offer promising alternatives, established techniques are primarily applied to highway anomaly detection and are rarely adapted for bicycle roads. Furthermore, most related studies lack clear definitions for anomalies and transparent annotation procedures, resulting in a shortage of effective datasets and detection methods specifically for bicycle pavement.
To address these issues, this research puts forward a new model, the Vector Quantized Transformer Autoencoder (VQTransAE), which integrates self-attention mechanisms with discrete latent representations to monitor bicycle road conditions using smartphone sensor data. Field experiments conducted on various pavement types, including asphalt, concrete, and brick surfaces, demonstrate the model's ability to accurately detect and locate anomalies such as cracks and potholes. Comprehensive testing using different smartphone platforms (Apple and Android devices) validates the method's robustness across differing sensor characteristics. A systematic comparative analysis between traditional bicycles and electric bicycles (e-bikes) also reveals a noticeable spatial overlap in their detection results. In addition, this study defines a standardized data processing workflow and establishes a rigorous manual annotation protocol, producing a publicly available, high-quality labeled dataset for future research.
However, current limitations include an insufficient representation of data from extreme seasonal conditions and unvalidated threshold generalizability across diverse cyclist characteristics. Future research directions include leveraging crowdsourced data collection for long-term infrastructure monitoring, optimizing anomaly thresholds to accommodate different cyclist conditions, integrating e-bike data to increase model coverage, and creating a quantitative link between anomaly scores and the degree of pavement deterioration.
Comments
Copyright 2025, Kewei Ren