Detecting Falls with Wearable Sensors Using Machine Learning Techniques

EasyChair Preprint 10154

Abstract

This study presents an innovative approach to the problem of fall detection, leveraging wearable sensor technology. Initially, we delineate the definition of falls, followed by an examination of their classification methods and categories. To address this issue, we propose a typology that employs both Long Short-Term Memory (LSTM) and a hybrid Convolutional Neural Network (CNN1D + LSTM). These models are trained to detect falls using data from accelerometers, gyroscopes, and magnetometers. Our proposed network models are trained and rigorously evaluated using a comprehensive wearable sensor dataset. They are also benchmarked against a range of different classifiers for comparative purposes. The LSTM model demonstrated an impressive accuracy of 98.04%, while the hybrid CNN1D + LSTM model achieved an exceptional accuracy of 99.68%. To further validate our approach, we compared the performance of our models against other deep neural network architectures that have previously been proposed and implemented. Our models demonstrated competitive, if not superior, performance, endorsing their potential for effective real-world application in fall detection.

Keyphrases: deep learning, fall detection, neural network

@booklet{EasyChair:10154,
  author    = {Shailendra Bhandari and Negar Elmisadr and Bereket Zerabruk Tekeste and Raju Shrestha},
  title     = {Detecting Falls with Wearable Sensors Using Machine Learning Techniques},
  howpublished = {EasyChair Preprint 10154},
  year      = {EasyChair, 2023}}