Human Activity Recognition Using Optimized Deep Learning with Data Augmentation

Abstract

Human Activity Recognition (HAR) involves classifying human movements and has become essential for assessing the frequency and length of different human acts. The development of intelligent assistive devices and the examination of manual operations both depend on this. HAR has recently made use of deep neural networks, especially when it comes to day-to-day tasks, utilizing multichannel time-series obtained from bodyworn devices equipped with various sensors. To begin, the WISDM and MHEALTH datasets are utilized as input, and a Generative Adversarial Network (GAN) is employed to learn a generative model that produces time-series data exhibiting similar space and time dependencies as real data. Subsequently, an Optimized Long Short-Term Memory Neural Network with Improved Aquila Optimizer (IAO-ILSTM-NN) is applied to perform HAR. The random parameters impact on prediction accuracy is addressed by optimizing the ILSTM-NN structure parameters using the IAO. Based on the simulation outcomes, the recommended framework executes superior than alternative frameworks that use the same datasets and baseline models. This underscores the effectiveness of the model in enhancing human activity recognition, particularly on multimodal sensing devices.