Human Activity Recognition-Oriented Incremental Learning with Knowledge Distillation

Recently, a variety of different machine learning methods improve the applicability of activity recognition systems in different scenarios. For many current activity recognition models, it is assumed that all data are prepared well in advance and the device has no storage space limitation. However, the process of the sensor data collection is dynamically changing over time, the activity category may be continuously increasing, and the device has limited storage space. Therefore, in this study, we propose a novel class incremental learning comprehensive solution towards activity recognition with knowledge distillation. Besides, we develop the representative sample selection method to select and update a specific number of preserved old samples. When new activity classes samples arrive, we only need the new classes samples and the representative old samples to preserve the network's performance for old classes while identifying the new class. Finally, we carry out experiments using two different public datasets, and they show good accuracy for old and new categories. Besides, the method can significantly reduce the space required to store old classes samples.