Model Personalization for Human Activity Recognition

Human activity recognition (HAR) has applications to various fields. However, not accounting for the personal differences among various subjects can lead to significant accuracy degradation. To address this problem, we propose a lightweight personalization process that enables an HAR model adapt for various users, some of them ever unseen before. Indeed, by adding a small amount of new labeled data, the model we propose can be personalized and boost the HAR accuracy when dealing with a wide range of target users. Furthermore, we also propose an innovative training algorithm to support personalization during the training stage. Our evaluations on three public real-world datasets demonstrate the superiority of our personalization approach, i.e. 6-14% improvement on the target-domain accuracy, while using only five labeled data points per-class.