Contextual movement models based on normalizing flows

Abstract Movement models predict positions of players (or objects in general) over time and are thus key to analyzing spatiotemporal data as it is often used sports analytics. Existing movement either designed from physical principles or entirely data-driven. However, the former suffers oversimplifications achieve feasible interpretable models, while latter relies on computationally costly, a current point view, nonparametric density estimations require maintaining multiple estimators, each responsible for different types movements (e.g., such velocities). In this paper, we propose unified contextual probabilistic model based normalizing flows. Our approach learns desired densities by directly optimizing likelihood maintains only single that can be conditioned auxiliary variables. Training simultaneously performed all observed movements, resulting an effective efficient model. We empirically evaluate our professional soccer. findings show outperforms state art being orders magnitude more with respect computation memory requirements.