Research on open set based class-incremental learning for human activity recognition

Through time series signals collected by wearable sensors, human activity recognition (HAR) needs to be carried out based on the given categories of training samples. However, there always exist new classes of data in real environment. These days how to effectively distinguish these new categories of data from the given classes become an important issue for HAR. Class-incremental learning aims to update the existing model using new knowledge when the target data is increasing. Also, the open set based recognition algorithm can provide the rejection option for the classifier to identify the target class that the model didn’t learn before. In this paper, an open set based class-incremental learning HAR framework is designed, which can continuously identify and learn new unknown classes. The framework combines extreme value model (EVM) with incremental learning to learn and recognizes new data. Here, PCA dimensionality reduction for features is applied to calculates the cosine, Euclidean and Manhattan distances between features, respectively. The simulation result reveals that the proposed model performs well on the UCI and PAMAP2 datasets compared with the existing open set based schemes. Higher accuracy can be achieved through PCA reduction with cosine distance calculation. Also, in the class incremental learning experiment, the proposed model can maintain high accuracy while new classes can be effectively identified.