Li, Joshua QiangAnsari, Fatemeh2025-10-132025-10-132025-07https://hdl.handle.net/20.500.14446/346828Ensuring the safety and durability of roadway infrastructure relies heavily on accurate pavement condition assessment and effective crash prediction. Conventional techniques for detecting pavement distress—such as manual inspections and contact-based friction tests—are resource-intensive, time-consuming, and susceptible to subjectivity. Additionally, traditional Safety Performance Functions (SPFs), which are generally derived from basic roadway and traffic characteristics, often fail to account for the influence of surface conditions on crash likelihood. Recent advancements in computer vision and deep learning have enabled the automation of pavement assessment through image-based analysis. Notably, the integration of multimodal data sources, including 2D grayscale images and 3D depth data, has shown promise in enhancing the detection of subtle surface irregularities under various environmental conditions. Deep learning (DL) architectures, including convolutional neural networks (CNNs) and mixed transformers (MiT), provide robust capabilities for pixel-level segmentation and feature extraction from complex pavement surfaces. Two different deep learning models have been developed in this study for multimodal multi-object pavement feature detection, including a U-Net-based model and a SegFormer-B5 model with MiT backbones. Both models demonstrate effective capabilities for identifying various pavement features, achieving overall accuracy of 97.10% and 98.24% for the U-Net-based model and the SegFormer-B5 model, respectively. Simultaneously, crash modeling can benefit from incorporating detailed pavement surface condition indicators, especially texture parameters that affect tire–road interaction and vehicle control. This study introduces a comprehensive framework that leverages high-resolution 3D imaging and DL to automate pavement condition assessment and safety evaluation at highway speeds. A mobile sensing platform equipped with a 0.1-mm resolution 3D laser safety sensor and a 1-mm resolution 3D condition survey system was deployed to collect both high-resolution data at low speed and lower-resolution data during highway-speed operation. A super-resolution model, PT-SRGAN based on a recursive generative adversarial network (GANs), was employed to reconstruct 0.1-mm texture surface data from 1-mm highway-speed imagery, enabling both macro- and micro-texture features to be captured for safety evaluation. Subsequently, texture metrics at both levels and seven 3D parameters (provide details) were calculated. Data from 27 representative pavement sections, along with historical crash data, were used to develop texture enhanced SPFs based on the reconstructed texture datasets. This study combines deep learning and high-resolution texture analysis to improve pavement evaluation and crash prediction. The U-Net and SegFormer-B5 models precisely identify surface features, while the PT-SRGAN model reconstructs detailed textures from highway-speed data. These metrics were then incorporated into the Safety Performance Functions, providing a scalable, data-driven framework for enhanced roadway safety management.application/pdfCopyright is held by the author who has granted the Oklahoma State University Library the non-exclusive right to share this material in its institutional repository. Contact Digital Library Services at lib-dls@okstate.edu or 405-744-9161 for the permission policy on the use, reproduction or distribution of this material.Thesisartificial intelligencepavement texturetransportation safety