🛡️ Fall Detection and Pose Classification

A hybrid AI system that leverages YOLOv8 and MediaPipe for real-time detection of falls, phone usage, and potentially unsafe worker behavior on industrial floors. Built for edge deployment on devices like the Raspberry Pi 3.

---

📌 Table of Contents

• Overview
• Approaches
• Hardware & Software
• Directory Structure
• Installation
• How to Run
• Results
• Challenges
• Future Work
• Contributors

---

🧠 Overview

This project aims to improve industrial safety by:

• Detecting fall incidents
• Identifying if a worker is using a phone
• Classifying poses like standing, sitting, falling, or sleeping

We achieve this by integrating:

• Object detection via YOLOv8
• Pose estimation via MediaPipe
• Custom logic to classify behavior based on pose landmarks

---

🔬 Approaches

⚙️ Approach 1 – Bounding Box Fall Detection

This approach relies on bounding box geometry during person detection:

• If a person’s bounding box suddenly shifts from vertical to horizontal, it's flagged as a potential fall.
• The width-to-height ratio is used to infer orientation.
• Limitation: Not reliable if person is partially visible or occluded.

🕺 Approach 2a – MediaPipe Pose Classification

This uses Google’s MediaPipe to extract 33 key body landmarks:

• Uses position and visibility of landmarks (e.g., shoulders, hips, nose).
• Custom rules classify posture:
  • Low head + no limb movement → Sleeping
  • Rapid vertical drop in hip/ankle Y → Fall
  • One hand near ear/head → Phone call
• Accurate in good lighting, but fails under occlusion.

🧍‍♂️ Approach 2b – YoloPose (Pose via YOLO)

YOLOPose integrates keypoint detection into YOLOv8 architecture:

• It predicts joint locations as an extension of person detection.
• Combines the speed of YOLO with pose estimation accuracy.
• Requires custom dataset and training setup.

📐 Approach 3 – Spine Vector Deviation

This novel technique computes the spine angle using:

• A vector from shoulder center to hip center.
• A tilt beyond a certain threshold implies a fall.
• More mathematically grounded and reliable for fall detection without relying on bounding boxes.

---

🧰 Hardware & Software

📦 Software

• Python 3.10+
• YOLOv8
• MediaPipe
• OpenCV
• NumPy
• Matplotlib

💻 Hardware

• Webcam (1080p)
• Raspberry Pi 3 (for edge deployment)

---

📁 Directory Structure

Fall-Detection-and-Pose-Classification/
│
├── Approach 1 - Bounding Box/
├── Approach 2a - Mediapipe/
├── Approach 2b - YoloPose/
├── Approach 3 - Spine Vector/
├── Output Images/
│
├── combined_yolo_&_mediapipe.py     # Main hybrid implementation
├── deploy_rpi3.py                   # Raspberry Pi deployment script
├── Implementation Report.docx       # Detailed technical report
└── README.md                        # This file

---

⚙️ Installation

Clone the repository:

git clone https://github.com/InvictusRex/Fall-Detection-and-Pose-Classification.git
cd Fall-Detection-and-Pose-Classification

Install dependencies:

pip install -r requirements.txt

---

🚀 How to Run

Run YOLOv8 + MediaPipe hybrid detection:

python combined_yolo_&_mediapipe.py

Deploy on Raspberry Pi:

python deploy_rpi3.py

---

📊 Results

• Detected falls using bounding box aspect ratio changes and spine vector tilt.
• Identified mobile usage using YOLOv8 object detection.
• Pose classification based on MediaPipe landmarks.
• Real-time performance tested on Raspberry Pi 3 (with optimization).

Sample output images are available in /Output Images.

---

⚠️ Challenges

• Lighting variation affects YOLO detection.
• MediaPipe fails in extreme body angles or occlusion.
• Limited dataset for edge cases (e.g., crawling, kneeling).
• Resource constraints on Raspberry Pi during concurrent model inference.

---

🔮 Future Work

• Improve pose classification with LSTM/Transformer-based temporal models.
• Extend detection to other safety gear (e.g., helmet, gloves).
• Alert system integration with MQTT or SMS.
• Deploy on NVIDIA Jetson Nano for better real-time performance.