Local approximation models for human physical activity classification

The research is devoted to the time series classification. The time series is measured by an accelerometer of a wearable device. A class of physical activity is defined by its feature description of a time segment. To construct this description, the authors propose to use parameters of various approximation splines (algebraic, smoothing, adaptive regression, or spline with dynamic nodes). The logistic regression is used as a classifier. It delivers desired quality of the activity recognition. The authors analyze the space of the local approximation parameters. Classification accuracy depends on the method of this space construction. The computational experiment finds the optimal approximation parameters and parameters of the classifier.