Human activity recognition (HAR) has become a focal point of research due to an increased usage of sensors and the rapid advancements in artificial intelligence (AI). With diverse applications ranging from monitoring the well-being of elders to crime prevention, the research in HAR spans various datasets and methodologies. Existing approaches have certain limitations, i.e., measurement equipment can be sizable and cause user discomfort. In real-world applicability, these approaches can create issues with data storage, and moreover raise privacy concerns for users unwilling to share sensitive information as active movements. In this paper, we conduct a comprehensive analysis of three distinct HAR approaches, each employing two machine learning (ML) algorithms. The first two use existing approaches for HAR and serve as a baseline for model performance. The third approach is a Federated Learning (FL) approach with classic ML algorithms, aimed at maintaining low data storage requirements and ensuring data privacy. The results obtained with the FL approach were found to be comparable to the classic ML methods, making the FL approach applicable in practice, whilst at the same time offering the benefits of FL.