Prediction of defensive player trajectories in NFL games with defender CNN-LSTM model

Defensive players in football have complex trajectory patterns that are hard to accurately model and evaluate. This is due to the defenders’ inherently reactive role, which requires them to dynamically respond to offensive strategies and on-the-spot decisions by the quarterback. Amongst defensive players, defensive backs are the most challenging to model. They often travel long distances during passing plays to fulfill their coverage responsibilities and change trajectories depending on their perception of where the ball will go. This is in contrast to other defensive players, whose total distance travelled and directions of movement are generally much more limited.  

There are 11 players lined up on each of the offensive and defensive side for every play in a National Football League (NFL) game, which translates to 22 players interacting with each other at any given time. Individual player trajectories are affected by their personal assignments, the current overall strategy (e.g., man versus zone coverage), and the movement and decisions of surrounding players and the ball. A football human expert is able to evaluate and predict a defensive back’s trajectory, as there exists an innate ability in humans to predict near-future events and take sequential actions while accounting for complex arrays of factors and potential outcomes via joint attention. However, evaluating large numbers of interacting inputs to generate sequential predictions remains difficult.

Developing the ability to predict and evaluate these trajectories is of paramount importance to better assessments of defensive coverage, offensive strategy, quarterback decision-making quality and even the probability of winning plays and games. It is even more challenging to predict “what-if” scenarios – for example, how should the defensive backs’ trajectories change if the receiver targeted in a passing play is changed? Answering these questions in a quantitative manner can provide talking points to football enthusiasts, and also help analytics-driven teams to better understand their offensive and defensive decisions...