Project Information

Annually, 37 million falls occur globally, posing a significant risk, especially for isolated elderly individuals.

As part of my semester project, I've developed a real-time fall detection system using Streamlit, Openpifpaf, and PyTorch.

This method, based on the research Multi-camera, multi-person, and real-time fall detection using long short term memory by Mohammad Taufeeque, Samad Koita, Nicolai Spicher, and Thomas M. Deserno. uses OpenPifPaf for pose estimation and LSTM models to classify falls into five classes: None, Normal, Falling, Warning, Fallen.

The five features include:

  • Aspect ratio of the bounding box, capturing body configuration.
  • Logarithm of the angle between the vertical axis and the torso vector, adding non-linearity.
  • Rotation energy based on torso and head vectors, measuring motion.
  • Derivative of the bounding box ratio, showing changes in body configuration.
  • Generalized force using joint angles and velocities over 3 frames, capturing motion acceleration.

With the potential to save lives, this fall detection system addresses a critical need for vulnerable groups of people.