Classification of Outdoor Sports Using Symbolic Fourier Transform of Multivariate Time Series

This study leverages a previously introduced multivariate time series dataset of sports activities, recorded from an individual athlete, to develop and evaluate advanced classification techniques. The dataset, consisting of activities such as walking, running, biking, skiing, and roller-skiing, captures three-dimensional features: heart rate, speed, and altitude. Utilizing this data, we applied the WEASEL+MUSE classification algorithm, renowned for its efficiency in handling multivariate time series, to distinguish among the activity categories based on their temporal characteristics. The classification was augmented using an early time series classification algorithm, which helped determine the optimal data usage for balancing accuracy and computational efficiency. Our results demonstrate high integrity and robust classification capabilities, achieving up to 93.0% accuracy. However, challenges were noted in distinguishing between closely similar activities, with an analysis indicating that using 33% of the data could still yield 85.6% accuracy in test scenarios. These findings underscore the potential of advanced machine learning models in enhancing activity recognition, particularly in complex, realworld conditions.