Latent Space Representation of Adversarial AutoEncoder for Human Activity Recognition: Application to a low-cost commercial force plate and inertial measurement units

Abstract

Human Activity Recognition (HAR) is a key component of a home rehabilitation system that provides real-time monitoring and personalized feedback. This research explores the application of Adversarial AutoEncoder (AAE) models for data dimensionality reduction in the context of HAR. Visualizing data in a lower-dimensional space is important to understand changes in motor control due to medical conditions or aging, to aid personalized interventions, and to ensure continuous benefits in remote rehabilitation settings. This makes patient assessment effective, easier, and faster.

In this study, the classification performance of the latent space created by the AAE is evaluated using the Wii Balance Board (WiiBB) and/or three Inertial Measurement Units (IMUs) placed on the forearms and hip. Various sensor configurations are considered, including only WiiBB, only IMUs, combinations of WiiBB with the IMU at the hip, and combinations of WiiBB with the 3 IMUs.

The accuracy of the latent space representation is compared with two common supervised classification models, which are the Convolutional Neural Network (CNN) and the neural network called CNNLSTM, which is composed of convolution layers followed by recurrent layers. The approach was demonstrated for two different sets of exercises consisting of upper and lower body exercises collected with 19 participants.

The results show that the latent space representation of the AAE achieves a strong classification accuracy performance while also serving as a visualization tool. This study is an initial demonstration of the potential of integrating WiiBB and IMU sensors for comprehensive activity recognition for upper and lower body movement analysis.