Enhancing latent space representation in Adversarial Autoencoders for exercise recognition: A data augmentation perspective using low-cost sensors

Exercise monitoring in the context of Human Activity Recognition (HAR) is essential for delivering immediate feedback and facilitating the analysis of movement. When combined with data dimensionality reduction, it can offer deeper insights into movement patterns, thereby aiding in the development of training programs. This study investigates the impact of data augmentation and the number of participants in the training data on the accuracy of 2D latent space representations generated by an Adversarial AutoEncoder (AAE). In this study, data from the Wii Balance Board (WiiBB) and Inertial Measurement Units (IMUs) placed on each forearm and hip were collected from 20 participants. Experiments were performed for upper and lower body exercises, with the accuracy of the latent space representation analyzed by varying the number of participants in the training set from 2 to 12 with and without data augmentation. The results demonstrate that the incorporation of data augmentation significantly improves the accuracy of the latent space representation of AAE. For example, using only two participants in the training set, data augmentation improves test accuracy by 10.2% and 4.4% for WiiBB data and IMU, respectively, for lower body exercises, while upper body exercises showed improvements of 6.8% and 3.1% respectively. These findings show how data augmentation can mitigate the limitations of small training datasets significantly improving latent space representations for HAR applications. This study emphasizes the importance of combining data augmentation strategies and sensor types to achieve reliable and interpretable results in remote rehabilitation systems.