Pose-Vision: Pose-Aware PPE Compliance (YOLOv8)

Real-time PPE compliance monitoring that fuses object detection (PPE) with human pose estimation to flag unsafe conditions on the fly. Works on webcam or video files, overlays a clean HUD, and can log per-frame events for later analysis.

TL;DR → Plug camera/video in → YOLOv8 (PPE + Pose) → rule engine → annotated video & JSON/CSV events.

---

Final Demo(screen recording)

example1_cut_processed.mp4

---

Prototype Demo(screen recording)

example_workers.1.mp4

---

✨ Features

• 🔍 Dual-model fusion: YOLOv8 PPE detection + YOLOv8 Pose keypoints
• ⏱ Per-frame compliance checks: Hardhat, Mask, Safety Vest + posture cues
• 🎥 HUD overlays: Bounding boxes, skeletons, compliance banners
• 📹 Webcam & file support: Stream live or process pre-recorded videos
• 🖥 Pure PyTorch: No OpenVINO required
• 🗂 Sample assets: Ready-to-test images & videos

---

🗂️ Repository Structure

pose-vision/
├─ scripts/
│  ├─ video_pipeline.py
│  ├─ webcam_pipeline.py
│  └─ test/
│     ├─ fine_tuned_posture.py
│     ├─ improved_pose.py
│     ├─ yolo_pose_image.py
│     ├─ yolo_pose_video.py
│     ├─ yolo_pose_webcam.py
│     └─ yolo_webcam.py
├─ models/
│  ├─ ppe.yaml
│  ├─ yolov8n-pose.pt
│  └─ yolov8n-ppe.pt
├─ assets/
│  ├─ images/input.jpg
│  └─ videos/
│     ├─ example1.mp4
│     ├─ example2.mp4
│     ├─ example_workers.mp4
│     ├─ output_pose.mp4
│     └─ SCreenRec.mp4
└─ Presentation/
   ├─ AI-Powered-PPE-Compliance-Monitoring.pptx.pdf
   └─ Pose-Aware-PPE-Compliance-System.pptx.pdf

---

🧠 Models & Classes

• Pose model: models/yolov8n-pose.pt
• PPE detector: models/yolov8n-ppe.pt
• Class map (ppe.yaml):

0: Hardhat
1: Mask
2: NO-Hardhat
3: NO-Mask
4: NO-Safety Vest
5: Person
6: Safety Cone
7: Safety Vest
8: machinery
9: vehicle

---

⚙️ Setup

Python 3.9+ recommended.

# Clone repo
git clone https://github.com/Harshitheccentric/pose-vision.git
cd pose-vision

# Create virtual environment
python -m venv .venv
# Windows: .venv\Scripts\activate
# Linux/Mac:
source .venv/bin/activate

# Install dependencies
pip install --upgrade pip
pip install ultralytics opencv-python numpy torch torchvision pyyaml tqdm

✅ No OpenVINO dependency needed.

---

🚀 Quickstart

Run on webcam

python scripts/webcam_pipeline.py \
  --ppe-weights models/yolov8n-ppe.pt \
  --pose-weights models/yolov8n-pose.pt \
  --source 0 \
  --view \
  --save

Run on a video file

python scripts/video_pipeline.py \
  --ppe-weights models/yolov8n-ppe.pt \
  --pose-weights models/yolov8n-pose.pt \
  --input assets/videos/example_workers.mp4 \
  --out runs/annotated_example.mp4 \
  --save-logs runs/example_events.json

---

🧩 How It Works

1. PPE detection – YOLOv8 finds workers + gear
2. Pose estimation – YOLOv8-Pose gives 17 keypoints
3. Fusion – Match PPE boxes with persons (IoU overlap)
4. Posture analysis – Angles & heuristics (from fine_tuned_posture.py)
5. Rule engine – Flag violations (e.g., no helmet, bent spine)
6. Output – Annotated video + optional JSON/CSV logs

---

🏗️ Architecture

flowchart LR
  A[Video or Webcam Source] --> B[Frame Grab]
  B --> C1[YOLOv8 PPE Detection]
  B --> C2[YOLOv8 Pose Estimation]
  C1 --> D[Fusion and Association person to PPE overlap]
  C2 --> D
  D --> E[Posture Module angles and heuristics]
  E --> F[Compliance Rule Engine]
  F --> G[HUD Overlay boxes skeletons banners]
  F --> H[Event Logger JSON or CSV]
  G --> I[Annotated Video Output]
  H --> I

Loading

---

🧪 Test Scripts

Located under scripts/test/:

• yolo_webcam.py → PPE-only webcam
• yolo_pose_webcam.py → Pose-only webcam
• yolo_pose_video.py → Pose on video
• yolo_pose_image.py → Pose on single image
• improved_pose.py → Enhanced visualization
• fine_tuned_posture.py → Posture classification logic

---

📊 Output Examples

Video: Annotated MP4 with skeletons & labels Logs (JSON/CSV):

{
  "frame": 123,
  "person_id": 5,
  "violations": ["NO-Hardhat", "NO-Safety Vest"],
  "posture": {"back_bend_deg": 32.1, "risk": "medium"},
  "timestamp": "00:00:04.10"
}

---

🔁 Training (Optional)

Train PPE detection

yolo detect train data=models/ppe.yaml model=yolov8n.pt imgsz=640 epochs=50 batch=16

Train custom pose model

yolo pose train data=your_pose.yaml model=yolov8n-pose.pt imgsz=640 epochs=100

---

🐛 Troubleshooting

• No webcam feed → Ensure --source 0 is correct and camera is free
• CUDA not detected → Install correct PyTorch build for your GPU
• Slow FPS → Use yolov8n for speed; scale up to s/m/l only if GPU allows

---