Symbolic Representation of Multivariate Time Series for Sport Activity Classification

Abstract

Abstract This study introduces a new multivariate time series (MTS) sport activity dataset consisting of five categories, walking, running, biking, skiing, and roller skiing. The original data of 228 activities have been recorded by an individual athlete for a 16 months time period in uncontrolled environments using two types of sport watches. The dataset consists of three-dimensional multivariate time series features such as heart rate, speed, and altitude, which are popular and pure sensor based attributes for endurance outdoor sport activities. The pre-processed signals were split into 69 seconds equal length segments and several segments from each single activity were gathered in order to conduct data augmentation because of the small dataset size. The MTS classifier called WEASEL + MUSE was applied to the dataset in order to discriminate categories based on the time series characteristics of the signals. The classification results was analyzed using several popular quality metrics and tools such as ROC curve. In addition, an early time series classification (eTSC) algorithm called TEASER was applied to determine how much data will be sufficient to find a balance in accuracy and computation time tradeoff. According to the results, dataset integrity is generally good and sport activities were classified fast and accurate, up to 93,0%. Signal length analysis indicated that 33% of the data will provide satisfactory results, 85,6% accuracy in the test data.