ORCID

https://orcid.org/0009-0001-4012-824X

Year

2023

Season

Fall

Paper Type

Master's Thesis

College

College of Computing, Engineering & Construction

Degree Name

Master of Science in Computer and Information Sciences (MS)

Department

Computing

NACO controlled Corporate Body

University of North Florida. School of Computing

First Advisor

Dr. Zornitza Prodanoff

Second Advisor

Dr. Sandeep Reddivari

Third Advisor

Dr. Thobias Sando

Department Chair

Dr. Asaithambi, Asai

College Dean

Dr. William Klostermeyer

Abstract

The traffic and roadway safety agencies spend significant efforts each year collecting roadway data, including lane configurations and other road surface marking data, such as areas with school zone markings, sidewalks, left turns, right turns, bicycle lanes, etc., for safety analysis and planning purposes. The current manual data collection methods pose significant operational and quality control challenges as they are costly and prone to errors. In addition to that the manual data collection is labor intensive and takes too much time involving high equipment costs, questionable data accuracy guarantees, and concerns about the safety of the crew.

This study aims to develop an automatic mechanism for extracting selected road surface marking features from the aerial images, an approach that will help the traffic and road safety agencies to cut down the cost of time, money, and safety of the crew by automating the data collection process.

To do that, a YOLOv5 deep-learning model has been built and trained to identify eleven roadway surface markings from the geo-referenced aerial images. The model showcases robust performance across various classes, demonstrating its versatility in identifying different types of roadway surface markings. For instance, when focusing on left-only classes, the model achieves a precision of 95% and 94% at a confidence threshold of 0.7. Similarly, for right-only classes, the model attains 93% precision and 92% recall at a confidence level of 0.6. Based on the literature reviews [1], the observed metrics are reasonably good and maybe improved further to allow the adoption of this mechanism by the traffic and safety agencies at scale.

Download

Share

COinS