Hyper parameter Optimization of Deep Learning Models for Fall Recognition using Tree-structured Parzen Estimator
Rana Abdul Rahman Lateef, Ayad R. Abbas · 2022
one of the topics related to Human Activity Recognition (HAR) is recognizing the fall of anelderly person. Recently, smartphone sensors were employed in the fall recognition system. However, manually tuning hyperparameters of deep models may be time-consuming compared with automatic tuning. In this paper, a Tree-structured Parzen Estimator (TPE) hyperparameter optimization algorithm was applied to one-dimensional Convolutional Neural Network (1D CNN) and Long Short Term-Memory (LSTM) models to automatically select the best influence hyperparameters, resulting in high classification accuracy. The suggested method was applied to the MobiAct public dataset and to the realistic dataset to distinguish fallactivity from daily living activity. The results reveal that suggested method achieved an accuracy of 99.4% with 1D CNN, 97.5% with LSTM, and 96.3% with real data. Thus, automatic tuning of the hyperparameters of the deep model based on the TPE search method reliably improved the classification accuracy of fall activities and reduced the efforts of selecting them manually.