🪖 Helmet Detection — YOLOv8

📁 Project Structure

ai project practical/
│
├── app.py                          # ✅ NEW — Gradio UI for Hugging Face deployment
├── requirements.txt                # ✅ NEW — All Python dependencies
├── README.md                       # ✅ NEW — This file
│
├── train.py                        # Train the YOLOv8 model
├── test.py                         # 🔧 FIXED — Test on a single image
├── webcam.py                       # 🔧 IMPROVED — Live webcam detection
│
├── yolov8n.pt                      # Base YOLOv8 nano weights (for training)
├── output.mp4                      # Output video from testvideo.py
│
├── Data/
│   ├── data.yaml                   # Dataset config (classes, paths)
│   ├── indianroadbike.mp4          # Sample input video
│   ├── testvideo.py                # 🔧 FIXED — Run detection on a video file
│   ├── train/images/               # Training images
│   ├── valid/images/               # Validation images
│   └── test/images/                # Test images (a.jpeg, aa.jpeg, ...)
│
└── runs/detect/train-2/
    └── weights/
        ├── best.pt                 # ✅ Best trained model weights
        └── last.pt                 # Last epoch weights

🧠 What the Model Detects

🐛 Bugs Fixed & Changes Made

1. test.py — Critical Logic Bug Fixed

• Added MODEL_PATH, IMAGE_PATH, CONF_THRESHOLD constants at top (easier to change)
• Passed conf=CONF_THRESHOLD to model call (filters weak detections at source)
• no_helmet now takes priority over helmet in final output (safety first)

2. webcam.py — Minor Improvements

• Added MODEL_PATH constant instead of hardcoded string
• Raised confidence from 0.25 → 0.35 (reduces false positives)
• Added show_conf=True (displays confidence score on screen)
• Added stream=True (memory-efficient for continuous live frames)

3. Data/testvideo.py — Path & Robustness Fixes

• Added MODEL_PATH, VIDEO_PATH, OUTPUT_PATH, CONF_THRESHOLD constants
• Fixed cap.get(3) / cap.get(4) → proper constants CAP_PROP_FRAME_WIDTH / CAP_PROP_FRAME_HEIGHT
• Added fps or 25 fallback (some cameras/files return 0 FPS — this prevents a crash)
• Added os.path.exists() check — raises a clear error if video file is missing
• Added verbose=False to model call — stops spammy per-frame console output
• Note: Run this file from the project root, not from inside the Data/ folder

4. app.py — NEW File for Hugging Face

• Upload any image of a bike rider
• Model runs YOLOv8 inference
• Returns annotated image with colour-coded bounding boxes:
  • 🟢 Green box = Helmet detected
  • 🔴 Red box = No helmet detected
  • ⚪ Grey box = Other class (bike, driver, etc.)
• Displays verdict: ✅ HELMET or ❌ NO HELMET
• Shows top confidence score (e.g. 87.3%)
• Safety-first logic: if any person without helmet is found → verdict = NO HELMET

5. requirements.txt — NEW File

ultralytics>=8.0.0
torch>=2.0.0
torchvision>=0.15.0
opencv-python-headless>=4.8.0
Pillow>=10.0.0
numpy>=1.24.0
gradio>=4.0.0

▶️ How to Run Locally

Single Image Test

Live Webcam

Video File

Gradio Web UI (local)

🚀 Deploy to Hugging Face

Step 1 — Copy the model to root

Step 2 — Create a Hugging Face Space

1. Go to huggingface.co/new-space
2. Choose Gradio as the SDK
3. Upload these 3 files:
   • app.py
   • best.pt
   • requirements.txt

Step 3 — Done!

🏋️ Training Details

📌 Important Notes

• All scripts must be run from the project root folder (where train.py lives)
• The trained model is at runs/detect/train-2/weights/best.pt
• For Hugging Face, rename/copy best.pt to the project root
• testvideo.py is inside Data/ but still uses paths relative to the project root