Using Artificial Intelligence to Detect Falls
Jonathan Sturdivant, Nicholas A. Morris, Tiara Hendricks, Gülüstan Doğan, Michel Heijnen · 2023
This work aims to apply both traditional machine learning approaches and deep neural networks in human activity recognition. A multi-modal approach is used to identify falls both in a frame as well as across a video. The models use camera data from a single position as well as three-axis accelerometer data to identify falls. This research aims to present possibilities for an easily implementable model using affordable data sources and limit the burden on healthcare staff by mitigating false-positive results. In our first experiment, the traditional machine learning models used returned an accuracy of approximately 98 percent and in our second experiment, the deep learning model had an accuracy of 89 percent but had more difficulty determining if the subject was classified as falling.