Semantic Template Recognition of Human Activities in Wearable Sensor Data Using Siamese Network

Human activity recognition plays a pivotal role in various fields, such as healthcare monitoring, smart environments, and human-computer interaction. In this study, we propose a novel approach for sensor-based human activity recognition.The key contribution of our work consists of first, defining activity representations we call “semantic templates”, which represent prototypical activity patterns of different human activity classes; second, designing and implementing a novel lightweight and versatile classifier for sensor-based HAR that leverages template matching through a deep-learning Siamese network. Through a series of rigorous experiments conducted on three distinct public datasets, we also demonstrate that the proposed approach yields enhanced performance in recognizing human activities when compared to a traditional deep multi-class classifier for resource-constrained devices. Furthermore, we showcase how our approach outperforms previous works by up to 20% in classifying previously unseen activities, paving the way for developing class-incremental continuous learning techniques.