Human Activity Recognition Using Federated Transfer Learning

Smartphones offer a promising platform for continual behavioural analysis. However, due to the diverse range of settings and situations, capturing individuals’ behavioural patterns in real-world scenarios can be challenging, given the potential variability in users’ actions and behaviours. Equally crucial to identifying various physical activities is considering the user’s behavioural context when modelling and understanding human activity in natural environments. This paper introduces a novel approach by integrating physical activities with human behavioural contexts, providing a new paradigm for context-aware human activity recognition. The primary aim of the current research in human activity recognition is to comprehend human behaviour through the interpretation of sensory data. The human activity recognition database was constructed using recordings from thirty participants engaging in daily activities while wearing smartphones attached to their waists with implanted inertial sensors. Each individual performed fundamental activities such as walking, stair climbing, sitting, standing, and sleeping. The system was developed using different classification algorithms, including CNN and Federated Learning. The experimental results demonstrate the system’s accuracy in recognizing activities, and the outcomes are compared and visualized in the form of a graph.