Model Update in Wearable Sensors Based Human Activity Recognition

In this article wearable sensors based human activity recognition is approached with a case where personal data collected has a high inner activity variety. With this kind of approach, the model adaptivity as well as update becomes more important issues for the activity recognition models. In authors' previous article it was shown that with this kind of data the personal models do not always outperform the user-independent models and a self-organising maps distance based approach was introduced to be used in personal and UI model fusion. In this work, the idea is developed further and it is shown that the same procedure can and should be used as predictive model as well as in data labeling also when updating the models in real-life. It is shown that compared to results of updating personal model our approach gives from 0.6 to 8.2 percentage units improvement in overall accuracy. In practice, just updating personal model with its given somewhat mislabeled data can in the worst case even drop the recognition accuracy by several percentage units which is an unwanted effect in real world applications.