| dc.description.abstract | Road safety remains a pressing global challenge, with over 1.3 million fatalities annually
attributed to traffic accidents, a significant portion of which stem from misinterpretation or
non-compliance with road signs. In Bangladesh, where over 4,000 deaths occur yearly due to
road incidents, the inconsistent maintenance and standardization of traffic signage exacerbate
this issue, particularly hindering the deployment of autonomous vehicles (AVs). This study
presents the development and evaluation of a real-time traffic sign detection system using the
YOLOv8 architecture, tailored to address the unique environmental and infrastructural
conditions of Bangladesh. The system was trained on the yolobew dataset, achieving a mean
average precision (mAP50-95) of 0.97266, a peak precision of 1.00 at maximum confidence,
and a recall of 0.99872 at an optimal confidence threshold of 0.773, demonstrating robust
performance across diverse sign classes. Comparative analysis with Faster R-CNN and
RetinaNet highlights YOLOv8’s superior real-time inference speed of approximately 30
frames per second, alongside its efficiency in resource-constrained settings, despite challenges
posed by class imbalance and occluded signs. The research further explores the influence of
environmental factors such as weather and lighting on detection accuracy, identifying areas for
improvement. Future directions include dataset expansion to mitigate class imbalance and
enhancements in model efficiency through lightweight architectures and hybrid approaches.
The implications of this work are profound, offering a foundation for integrating the system
into AVs to enhance navigation safety, reduce human error, and support policy initiatives for
standardized signage, thereby contributing to a significant reduction in road accident rates in
Bangladesh as of August 2025. | en_US |
