Time Series Augmentation with Time-Scale Modifications and Piecewise Aggregate Approximation for Human Action Recognition

In this paper, a method for time series augmentation, aiming at the improvement of human action recognition accuracy of a deep learning classifier, is proposed. The approach performs time-scale modifications of the input time series and transforms them into compact sequences of time segments using Piecewise Aggregate Approximation (PAA) to facilitate the training of a neural network. The approach is compared against related methods on six representative datasets using Bidirectional Long Short-Term Memory (BiLSTM) classifier. It is shown that the resulting artificial time series lead to a better performance of the deep learning model than augmented data samples generated by popular approaches. The source code of the method is available at https://marosz.kia.prz.edu.pl/Adder.html.