Deep Learning-Based Stacking Neural Network and Generative Adversarial Networks for Human Activity Recognition Based on Ambient Sensors
Abstract
Smart home for healthcare services has acquired more attention since the increasing development of the Internet of Things and the population ageing over the world. Human activity recognition (HAR) is one of the concerns of the smart home. Ambient sensors based HAR is one promising direction. This research proposes a deep learning-based stacking method for HAR using ambient sensors. We first generate base models of convolutional neural networks (CNNs) and long short-term memory (LSTM) with different architectures, training data samples, and sliding window sizes. These base models are further integrated by a LSTM model to make final predictions. Furthermore, we propose a generative adversarial network to generate synthetic data as supplementary training data to tackle the problem of insufficient data. These two methods are used together on six real-world datasets. Results show that our proposed methodology statistically outperforms other approaches in the literature.