Title: Dynamic modeling of player movement in American football
Authors: Karl Pazdernik - Deep Football (United States) [presenting]
Jacques Kvam - Deep Football (United States)
Abstract: American football is a game of inches. The offense attempts to gain separation and yardage, while the defense works to collectively and continuously limit these distances. To truly understand the complex spatiotemporal patterns of offensive separation, identification of defensive coverage is necessary. The aim is to outline a novel methodology used to estimate the probability that each defender is tracking each offensive player at predetermined intervals of time within a play using a hidden Markov model. Within each defensive assignment, coverage types also exist. A defender may be in attack mode, in more of a surveillance motion, or may struggle to maintain proper coverage, trailing their assignment. We use a secondary group of hidden states to differentiate between these three behavioral patterns. From these estimated probabilities, unique summary statistics are possible. We can now quantify previously unmeasured statistics such as the amount of attention an offense player receives, the amount of separation a route runner can obtain, a defender's instincts regarding to their ability to diagnose a play, a defender's ability to recover when beaten, and the degree to which a defense attacks the ball carrier. For illustration, both simulated data and data from NFL games obtained through the All-22 game film are used.