An accelerometer based fall detection system using Deep Neural Network
Sankalp Garg, Bijaya Ketan Panigrahi, Deepak Joshi · 2019
Fall detection has integral role in elderly health care. Inertial sensors are popular choices to develop fall detection algorithms due to their portability, wearability, and no privacy concerns. Previous algorithms have used acceleration based features for fall detection; however such algorithms require feature engineering skills. In this paper, we propose a Deep Neural Network(DNN) for fall detection. A 5 seconds waveform, unfiltered and low-pass filtered with 10 Hz and 20 Hz, is fed to DNN for fall detection. The model is validated on two different datasets available online. The results report 95.1% classification accuracy, 94.4% sensitivity, 95.5% specificity for a dataset and 86% classification accuracy, 75% sensitivity, 92.1% specificity for other dataset. For the feasibility in real time application, the algorithm was implemented on an android based mobile system. The average computation time was noted to be 50 milliseconds. Finally, the algorithm proved to be independent of filtering operations (p = 0.551) suggesting the approach to be useful in noisy environment.