Feature generation for physical activity classification

The paper investigates human physical activity classification problem. Time series obtained from accelerometer of a wearable device produce a dataset. Due to the high dimension of object description and low computational resources one has to state a feature generation problem. The authors propose to use the parameters of the local approximation models as informative features. The experiment is conducted on two datasets for human activity recognition using accelerometer: WISDM and USC-HAD. It compares several superpositions of various generation methods and classification models.