Federated learning (FL) is an emerging technique for training machine learning models in a distributed manner and offers advantages in terms of data privacy and saving communication costs. Wearable devices, such as smart watches, are personal devices and generate lots of personal sensor data. In this respect, FL can offer advantages; hence, their data requires more privacy, and saving communication costs is essential. In this paper, we investigate the performance of FL compared to centralized learning for the domain of human activity recognition (HAR) using wearable devices. We used a dataset from the literature composed of sensor data collected from smart watches and trained three different deep learning algorithms. We compare the performance of the centralized and FL models in terms of model accuracy and observe that FL performs equivalent to the centralized approach.