Advancing spatiotemporal pattern analysis using top-level sports data
ASPAS
Abstract
The vast amount of collected human spatiotemporal movement data hides huge opportunities for various stakeholders and scenarios (e.g., fire evacuation, autonomous vehicles, infrastructure planning, and sports performance analysis). To reveal hidden insights from this data, one can use various state-of-the-art machine learning (ML) approaches. Our focus is to develop and evaluate new approaches for solving typical computational problems, which arise when analyzing huge amounts of spatiotemporal data such as detecting variable-length patterns and finding the right representation models for movement trajectories. As a testing bed, we use sports data, in this case a high-quality dataset from the German soccer Bundesliga. We choose semi-supervised learning (SSL) as a promising approach to exploit the unlabeled dataset to learn a compressed, latent representation of the data. Then, this embedding can be used as an encoding step to help with clustering and predicting movement patterns. The techniques developed and validated on our high-grade data provide a baseline for further research in all problems with pedestrian trajectory data.