Automated play event detection from sparse GPS American football data using recurrent neural networks

Contextualizing in-game actions from sparse GPS tracking data in American football is crucial for performance analysis but currently relies on time-consuming manual annotation. Existing automated play recognition tools primarily utilize video, leaving a significant gap for GPS-based solutions, especially with a variable number of tracked players. We address this by proposing an automated and flexible neural network approach for play event detection from sparse GPS data. We tested logistic regression, XGBoost, and two recurrent neural network models (GRU, LSTM) on 5020 data windows from 36 games, evaluating them against 962 play events in 8 hold-out games. Our recurrent neural network models achieved the best performance, with a mean F1-score of 89%, and maintained similar accuracy across diverse play types (passes, rushes, punts) and varying levels of on-field player coverage. Beyond event detection, we demonstrate the utility of the model for downstream applications, including accurate estimation of physical workload metrics and the contextual filtering of head impact events. This method drastically reduces event annotation time from hours to mere seconds per game. The developed play event detection model is publicly available, offering a valuable resource for researchers and practitioners to enhance sport performance contextualization in football.