Towards Intelligent Dietary Monitoring: Deep Learning Techniques for Alimentary Activity Detection

Human activity recognition (HAR) has become paramount in healthcare, well-being, and daily activity monitoring applications. Identifying activities such as eating and drinking can provide valuable insights into dietary habits and the management of chronic diseases. However, existing HAR models often disregard minority classes, particularly in recognizing subtle and underrepresented alimentary activities, which are often overlooked and receive less emphasis in studies. In this context, we propose a deep-learning architecture for recognizing human alimentary activities based on smartwatch sensor data. It integrates convolutional layers for feature extraction, Bidirectional Gated Recurrent Unit layers to capture long-term dependencies in both directions, and multi-head attention mechanisms to emphasize the most relevant features. We also include residual connections to ensure feature retention, improving classification performance. Experimental results show that our method obtained the best classification output, with a balanced accuracy of 86.55%, outperforming prior-art models. Our findings indicate that our approach enriches the recognition of complex alimentary activities, making it highly suitable for applications in healthcare and dietary monitoring.