This project aims to enhance mobility and independence for visually impaired individuals by assisting them in navigating real-world environments safely. The system uses YOLOv8s (with transfer learning) to detect obstacles in real-time and provides audio navigation instructions, enabling hands-free guidance.
- Real-time object detection on live camera feed
- Detects common campus obstacles:
- bike, car, cycle, light post, objects, pathhole, scooty, stairs, tree
- Voice-based navigation feedback (e.g., "Move Right", "Stop", "Turn Left")
- Works in dynamic outdoor and indoor environments
- Lightweight YOLOv8s model ⇒ fast + optimized for real-time
- Can be deployed on laptops, Raspberry Pi, Jetson, or mobile (future)
- Custom dataset captured at ABV-IIITM Gwalior campus
- Total images: 670
- Annotated using Roboflow
- Preprocessed and augmented (rotation, brightness, contrast, flip)
- Resized to 512 × 512
- Train/Val/Test split:
70% / 20% / 10%
- Model: YOLOv8s
- Training: Transfer Learning from pretrained COCO weights
- Framework: Ultralytics YOLO + OpenCV
- Parameters: 11.1M
- Compute Cost: 28.7 GFLOPs
| Metric | Score |
|---|---|
| Precision | 87.6% |
| Recall | 82.9% |
| F1-Score | 85.2% |
| mAP (0.50-0.95) | 55.8% |
Abhinandan Mandal