Tools
Tools
Ali, Hasan (2024) Street Navigation for Visual Impairment using CNN and Transformer Models. Masters thesis, Dublin, National College of Ireland.
Preview |
PDF (Master of Science)
Download (1MB) | Preview |
Preview |
PDF (Configuration Manual)
Download (10MB) | Preview |
Abstract
This paper addresses the challenge of street navigation for individuals with visual impairments and explores the potential of Artificial Intelligence (AI) to enhance navigation safety and effectiveness. We evaluate the performance of state-ofthe-art Computer Vision Object Detection models, focusing on accuracy and speed. The central question is whether Transformer-based Object Detection models outperform other models. We use the specialized dataset ”Walking On The Road” adapted to include only relevant classes, to compare deep learning and Transformer models in pre-trained and fine-tuned states. Metrics used include Mean Average Precision (mAP) for accuracy and Average Inference Time in milliseconds for speed. Our results show that YOLO models surpass Transformer-based models in both accuracy and speed. In Phase 1, YOLOv8x achieved the highest mAP of 0.399 with an average inference time of 14ms, while Transformer-based DETR had a lower mAP of 0.344 and a significantly longer inference time of 818.2ms. In Phase 2, after fine-tuning, YOLOv8x again outperformed with an mAP of 0.471 compared to DETR’s 0.323. These findings indicate that YOLO models are more effective for street navigation applications, providing superior accuracy and speed for visually impaired individuals.
Actions (login required)
 |
View Item |