Graduation Semester and Year
Summer 2026
Language
English
Document Type
Thesis
Degree Name
Master of Science in Civil Engineering
Department
Civil Engineering
First Advisor
Dr. MD Sahadat Hossain
Abstract
Plastic-modified asphalt pavements provide a potential approach to improving material sustainability while reducing plastic waste. However, their long-term field performance remains less documented than laboratory-based findings. This study developed a deep learning-enabled computer vision framework for detecting, segmenting, quantifying, and assessing surface cracking on a plastic-modified asphalt pavement section located near SH 205 in Rockwall County, Texas.
A YOLOv8n-seg instance-segmentation model was developed using publicly available pavement crack images that were curated and re-annotated with polygon-based masks. Model sensitivity to dataset size was evaluated using six incremental datasets containing 200, 400, 600, 800, 1,087, and 1,293 images. The models were trained using an 80:20 training-validation split for 200 epochs at a resolution of 600 × 600 pixels. The 1,293-image model was selected for field application based on its overall performance and stability, achieving a best mAP@0.50 of approximately 0.783 and a normalized true-positive proportion of approximately 0.85.
Field imagery was collected using a rear-mounted GoPro HERO 10 along the 0.66-mile monitored section. Separate median-side and shoulder-side videos were processed into chainage-referenced frames, resulting in 2,532 field images. Of these, 314 images from the initial 0.1-mile section were used for detailed detection and quantification, while 2,218 images from the remaining 0.2- to 0.66-mile section were used for extended screening. Among 17 manually confirmed crack frames, the model detected 14, corresponding to an 86% frame-level detection rate and a 17.6% missed-detection rate. Little to no clearly detectable cracking was observed in the remaining roadway section.
The 14 verified crack masks were processed through binarization and skeletonization to calculate segmented area, skeleton length, crack density, and geometric response ratio. These parameters provided a quantitative description of crack extent and morphology beyond visual detection alone. The resulting pavement-condition assessment produced a Pavement Condition Index of approximately 99.7, indicating excellent surface condition. The developed framework demonstrates the feasibility of integrating low-cost field imaging, instance segmentation, crack quantification, and condition assessment to establish an early-service baseline for future monitoring of plastic-modified asphalt pavement.
Keywords
Plastic-Modified Asphalt Pavement; Pavement Crack Detection; YOLOv8n-seg; Instance segmentation; Crack Quantification; Pavement Sustainability
Disciplines
Civil Engineering | Geotechnical Engineering
License
This work is licensed under a Creative Commons Attribution-NonCommercial-No Derivative Works 4.0 International License.
Recommended Citation
Maharjan, Garima, "DEVELOPMENT OF A DEEP LEARNING-ENABLED COMPUTER VISION FRAMEWORK FOR PAVEMENT CRACK ASSESSMENT ON PLASTIC-MODIFIED ASPHALT SECTION: SH 205, DALLAS DISTRICT" (2026). Civil Engineering Theses. 5.
https://mavmatrix.uta.edu/civilengineering_theses2/5