Sport Activity Classification Using Interlaced Multivariate Time Series Signals

This study investigates performance of Time Series Classification (TSC) models in non-laboratory conditions created dataset which contains sport activities in three categories, such as biking, running, and other. The main challenge conducted is to convert multivariate data to univariate. There are several methods to conduct that transformation, but in this case, we develop feature interlacement wherein three features are combined into the same signal from three dimensional multivariate time series data for each sport activity file, producing so called interlaced multivariate signals. Five univariate TSC models from sktime API evaluated in analysis are Time Series Forest, Supervised Time Series Forest, Random Interval Spectral Forest, Random Interval, and Shapelet Transform classifiers. Interlaced multivariate signal data was successfully constructed from the raw dataset and applied to TSC models achieving 91% accuracies on average among the models. Based on the overall scores obtained, applied data transformation method is well applicable in retrospective sport activity classification (SAC) using univariate time series analysis algorithms.