Optimising Deep Learning for Wearable Sensor-Based Fall Detection

Fall detection is essential for enabling prompt aid following such accidents. Despite the innovation in deep learning for Human Activity Recognition (HAR), previous studies have not identified a fundamental methodological setup of deep learning methods for fall detection properly. Therefore, this research delves into the deep learning methods for fall detection through an empirical study, employing fundamental deep learning techniques of HAR such as Temporal Convolutional Network, Multi-Channel Convolutional Network, etc. and two public datasets, CMDFall and UP-Fall datasets. We explore the influence of flattening and augmentation methods on various single-branch and multi-branch deep learning models within the HAR field and also explore feature combinations of accelerometer data. In order to identify the best setup for fall detection, comprehensive experiments were carried out spanning different parameters, probing the effectiveness of diverse flattening methods, data augmentation, feature engineering, and model architectures. The findings show how to improve the model’s performance with an optimal setting.