Fall Detection using Pose Estimation and GRU

Description

This project implements a fall detection system using pose estimation techniques and recurrent neural networks (GRU). The system is capable of analyzing video sequences to identify falls in real-time.

Project Structure

• notebooks/: Contains Jupyter notebooks for model training and evaluation.
• src/: Contains the project's source code.
  • models/: Implementations of fall detection models.
    • fall_detection_lstm.py: LSTM model for fall detection.
    • fall_detection_gru.py: GRU model for fall detection.
  • utils/: Utilities and helper functions.
    • video_detect_falls.py: Functions for fall detection in videos.
    • body.py: Definitions of connections and body parts for pose estimation.
• data/: Directory to store training and test data.
• media/: Contains images and videos used in the README and other documents.

Installation

1. Clone the repository:

   git clone https://github.com/your-username/fall-detection.git
   cd fall-detection

2. Create a virtual environment and install the dependencies:

   python -m venv .venv
   source .venv/bin/activate  # On Windows: .venv\Scripts\activate
   pip install -r requirements.txt

Usage

Fall Detection in Video

To detect falls in a video, use the following script:

from src.utils.video_detect_falls import video_detect_falls

video_detect_falls(
    video_path='data/videos/falls/yoga-fail-fall.mp4', # Change this to the path of the video you want to test
    yolo_model_path='yolo11x-pose.pt',
    gru_model='models/gru_model.pth',
    fall_threshold=.95,
    scale_percent=100,
    sequence_length=20,
    show_pose=True,
    record=True,
)