Human activity recognition is a key area in human–computer interaction because it enables systems to understand people’s actions and respond through implicit interaction, without requiring direct commands. This article presents a comparative analysis of six widely used public datasets, considering factors such as participant diversity, variety of activities, types of sensors, and capture conditions. The results show that although these datasets have driven the development of more accurate models, they still present limitations related to the lack of diversity and class imbalance, which affects the generalization ability of systems in real-world contexts. The need to design more inclusive and representative datasets that reflect the complexity of human interaction and support the development of fairer, more robust, and more adaptive human activity recognition systems is highlighted.
Mercado et al. (Sun,) studied this question.