Outdoor navigation for visually impaired people using YOLOv5 and Transfer learning: an analytical study

This study introduces an obstacle detection system for visual impairment rehabilitation in visually impaired individuals, leveraging YOLOv5 and transfer learning. The methodology comprises four main phases. First phase is sidewalk dataset generation, involving creating our customized PESID dataset and a proprietary dataset. Second phase is image annotation into eight essential obstacle classes. We pay special attention to edges, slopes, and holes—often overlooked in existing sidewalk datasets. These aspects are crucial for ensuring the safety of visually impaired individuals. Third phase is model pretraining with the large-scale COCO dataset. Last phase is fine-tuning using transfer learning on specialized datasets. The evaluated models for obstacle detection include YOLOv3, YOLOv5s, YOLOv5l, YOLOv5x, YOLOv5n, MobilenetSSDv2, Detectron2, RetinaNet, and Faster RCNN. On the customized PESID dataset, YOLOv5s, YOLOv5x, YOLOv5m, and YOLOv5l outperform other models in terms of Mean Average Precision at an Intersection of Union (IOU) of 0.5 (MAP0.5) and detection speed. On our proprietary dataset, the YOLOv5l model achieves a MAP0.5 of 84.5% and 77 frames per second. Notably, the MAP0.5 scores for edges, slopes, and holes are 77%, 85%, and 84%, respectively. Additionally, improving image annotation quality significantly enhances model accuracy by more than 10% for two datasets.