HAR-STGAN: Human Activity Recognition-Spatial Transformer network for Generative Adversarial Network
Aixin Nian, Songping Huang, Fei Wang, Yu Zhao · 2024
In this study, we tackle the critical challenge of imbalanced human acivity datasets, which has long hindered the advancement of human activity recognition (HAR) systems. To address this issue, we propose a novel multi-sensor data synthesis method based on generative adversarial networks (GANs). This innovative approach allows us to synthetically generate high-quality data for underrepresented behaviors, thereby balancing the dataset and enhancing the performance of downstream HAR tasks. Recognizing the importance of preserving the inherent relationships between sensor modalities, our algorithm considers the common features shared across multiple sensors. Through a spatial transformation mechanism, we are able to seamlessly translate the characteristics of different wearable sensors, ensuring the generated data maintains a high degree of consistency and specificity. To further bolster the authenticity and diversity of the synthetic data, we incorporate an auxiliary classifier that discriminates between real and generated samples. The classification loss is then seamlessly integrated into the GAN's training objective, providing valuable guidance for the network's optimization. Extensive evaluations across three key dimensions - temporal dynamics, data similarity, and category differentiation - demonstrate the HAR-STGAN algorithm's superior performance compared to state-of-the-art alternatives.