Fall Detection System Based on Simple Threshold Method and Long Short-Term Memory: Comparison with Hidden Markov Model and Extraction of Optimal Parameters

In an aging global society, a few complex problems have been occurring due to falls among the increasing elderly population. Therefore, are detected using pendant-type sensor that can be worn comfortably for fall detection. The sensed data processed by embedded environment and classified long-term memory (LSTM). A detection system combines simple threshold method (STM) LSTM, STM-LSTM-based system, is introduced. terms of training accuracy, proposed compared with previously reported STM-hidden Markov model (HMM)-based system. accuracy STM-LSTM 100%, while highest STM-HMM-based one 99.5%, which 0.5% less than best addition, in optimized LSTM this may overfitted because all trained without separating any validation data. order resolve possible overfitting issue, evaluated separately 4:1, then optimal values parameters normalization found as follows: 98.21% at no-normalization, no-sampling, 128hidden layer nodes, regularization rate 0.015. It also observed number hidden nodes or sampling interval increases, value increases. This means suppressed regularization, thus appropriate must selected improve efficiency.