Publication Date
2024
Document Type
Thesis
Committee Members
Fathi Amsaad, Ph.D. (Advisor); Wen Zhang, Ph.D. (Committee Member); Daniel Koranek, Ph.D. (Committee Member); Kenneth Hopkinson, Ph.D. (Committee Member)
Degree Name
Master of Science (MS)
Abstract
Traffic surveillance and enforcement heavily depend on the real-time detection of helmets and license plates, particularly in high-density urban environments. This study presents a dynamic and optimized lightweight model, the proposed G-YOLOv8n, designed for resource constrained edge devices like the Raspberry Pi. By integrating the GhostNet module into the YOLOv8n architecture, this research achieves a nearly 50% reduction in model size and computational load, while maintaining comparable detection accuracy to the original YOLOv8n. These enhancements enable real-time processing capabilities crucial for traffic monitoring operations. The growing demand for real-time, low-power solutions in intelligent transportation systems necessitates lightweight, efficient detection models. The proposed G-YOLOv8n model developed in this research leverages advancements in deep learning model compression and efficient feature extraction techniques to meet the constraints of edge computing platforms. Through methodologies of pruning and quantization, the model’s deployment on Raspberry Pi devices is optimized to balance detection accuracy with processing speed, addressing the challenges of limited memory and computational power typical of edge environments. In addition to object detection, this study also integrates Optical Character Recognition (OCR) for license plate recognition, enabling an integrated solution for helmet detection and automatic license plate recognition (ALPR) in traffic enforcement scenarios. Extensive experimental validation on real-world traffic datasets demonstrates the proposed G-YOLOv8n model’s ability to perform reliably under varied lighting and weather conditions. To ensure robustness, the model was subjected to adversarial patch attacks to evaluate its robustness under challenging conditions. While a strong indicator of performance in detecting license plates and riders is demonstrated by the model, it exhibited a degree of vulnerability to occlusions, particularly in the helmet and no-helmet classes. This research advances the field of real-time traffic surveillance by showcasing a scalable and energy-efficient approach for deploying advanced object detection on low-power devices. The findings offer valuable insights into model optimization strategies for edge deployment and adversarial attack resilience, paving the way for enhanced safety and compliance in intelligent transportation systems
Page Count
127
Department or Program
Department of Computer Science and Engineering
Year Degree Awarded
2024
Copyright
Copyright 2024, all rights reserved. My ETD will be available under the "Fair Use" terms of copyright law.
ORCID ID
0009-0004-8911-0851
