Vision-Based Fall Detection and Alarm System for Older Adults in the Family Environment

Abstract

This study proposes an innovative fall detection and alarm system for the elderly in the family environment based on deep learning. The overall cost of hardware development is a camera and an edge device like a Raspberry PI or an old laptop that can detect and alert users to falls without touching the user’s body. The development idea of the system is as follows: 1. Collect the pictures of falling and normal states under different conditions; 2. The improved lightweight SSD-Mobilenet object detection model is used to train the data set and select the optimal weight; 3. Optimal results are deployed on a Raspberry PI 4B device using a lightweight inference engine Paddle Lite. The mean Average Precision of the best model is 92.7%, and the detection speed can reach 14FPS (Frames Per Second) on the development board. When the camera detects that someone has fallen for 10 s, the compiled script sends an alert signal to the default guardian’s email via the Mutt email program on Linux. The experimental results show that the fall detection system achieves satisfactory detection accuracy and comfort.