Paying Attention to Human Activities: Recognizing Daily Activities with an Attention-based Neural Network and Inertial Sensors

Human Activity Recognition (HAR) is essential for a variety of real-life applications, especially within the Internet of Medical and Health Things (IoMT) where precise monitoring of complex human movements and postures is required. HAR technology plays a critical role in fields including rehabilitation therapy, sports performance analysis and training, occupational safety, chronic disease management, and supporting independent living for senior populations. Although artificial neural networks provide strong classification accuracy, their performance can be further enhanced by incorporating an attention mechanism, which enables the model to prioritize and focus on the most relevant data segments. In this paper, we introduce a novel deep learning model, AktionFocusNet, which integrates an efficient attention mechanism to advance HAR systems in the IoMT domain. AktionFocusNet utilizes multi-head attention to process multiple sections of the input simultaneously, with each attention head autonomously identifying specific features associated with the recognized activity. These outputs are then concatenated to form a detailed representation of the input data, allowing the model to capture a wide range of complex movement patterns and thus improve classification accuracy. We also reduce model complexity by calculating the magnitude of 3-D accelerations as input features. We evaluate the performance of AktionFocusNet on the WISDM Activity Prediction (WISDM 11) benchmark dataset, achieving an average accuracy of $\mathbf{9 7. 9 1 \%}$, and calculate individual $\mathbf{F 1}$ scores, recall, and precision for each class, demonstrating that AktionFocusNet outperforms existing models.