Football is a game of inches. You need to fight for every yard. Assuming that these common axioms are true, it is clear that gaining or, conversely, limiting space are paramount to the success of an offense or defense in American football. With the emergence of player tracking data in the NFL, appropriate analysis of this data is necessary to extract the maximum information. To meet this demand, we present a player tracking algorithm that estimates the probability that each defender is following each non-lineman offensive player for every play. Benefits of using this methodology include:
- Improved measure of separation in a route.
- Quantify how much attention a defense pays to any offensive player.
- Measure how quickly a defensive player diagnoses a play.
- Summarize a defense’s hustle in how it swarms to the ball carrier.
- Estimate how often a defender is beaten in pass coverage and how quickly they can recover.
- Differentiate a defender’s style of play between attacking, surveillance, and trailing.
- And more!
To estimate which defender is following which offensive player, we must consider both location and trajectory. Consider the plots below. Clearly, defender B is closer to the offensive player, but that doesn’t guarantee that he is trying to cover that player. Once you include trajectory, you see that in the plot on the left, defender A is tracking that offensive player and in the plot on the right, defender B is tracking that offensive player.
Not only do we consider who is tracking whom, but also how that defender is in pursuit. We consider three possible movement types: attack, surveillance, and trailing. We begin with the first two motion types. These can be interpreted as whether a defender is going directly towards the offensive player (assuming an appropriate angle of pursuit) or whether he is trying to shadow the offensive player. This distinction is illustrated in the left plot below.
These two motion types help describe the majority of motion on the football field. Even more complex motion, such as a pass rush moves and stunts along the defensive line, could reasonably be categorized as one of these motion types. However, the attacking and surveillance motion assume that the defender could possibly intersect the offensive player’s path at some point before he scores. Thus, another motion type is required to describe defenders who have been beaten. We define this as the “trailing” motion, illustrated in the right plot below.
The result is a probability of one of 18 categories for each defensive player for every frame of every play. The 18 categories are all combinations between the 3 motion types and the 6 eligible receivers on offense. Since the 5 set offensive lineman should not need to be guarded, even if a defender is moving toward one of these players, it is assumed that they are simply trying to get around them to tackle someone else. A simulated example play is provided below, with tracking probabilities given as pie charts.
Although the assignments in this play are fairly obvious, you can see that the free safety displays some hesitation before committing to the slot receiver and that the slot cornerback has a moment when he is trailing. The left defensive end, despite heading towards the running back, maintains the highest probability of attacking the quarterback for the majority of the play.
Given an estimated tracking probability for each defender, some interesting summary statistics are now possible that quantify never before measured aspects of football play.
When a positive play occurs, the credit is typically given to those directly involved in the play. However, if the eventual receiver is wide open because he beat single coverage, often part of the reason the play worked is because another receiver is demanding a double team. Although the boxscore will only give credit to the player that caught the ball, both players contributed. In fact, the player receiving the extra attention is typically the more coveted player. Those players who put up big boxscore numbers despite the extra attention are what we consider the superstar, high-impact players. So, in order to differentiate between hard-earned versus easily achieved statistics, we estimate the average number of defenders concentrated on an offensive player for a given play by averaging their tracking probabilities.
For example, in the play above, the defense sends a 5-man blitz, so the quarterback is receiving the most attention. The slot receiver gets the double team and this new metric captures that information.
Also included in the previous plot is the swarming defense metric. This statistic is computed in a similar fashion, the average of tracking probabilities, however with a different purpose in mind. Good defenses are known to “swarm” to the ball carrier. Translation: once the ball carrier/receiver is determined, a high percentage of the defenders will aggressively move towards the ball. We compute the average attention received by each offense player, but only after the eventual ball carrier is determined. This play results in a sack, so the eventual ball-carrier is the quarterback. But on running plays or successful passes, we can quantify defensive hustle at the team level using this statistic.
Some defenders do not have elite combine measurements, yet are always around the football and collect impressive statistics. Shortcomings in physical ability can be compensated by a sharp mind, able to quickly diagnose the play and identify the primary target. The tracking probability helps us understand this behavior. We can measure when a defender commits to guarding any offensive player, relative to the time of pass or handoff. Knowing the outcome of the play, we can highlight those players who made the right decision and how soon that decision was made. This helps measure defenders who are more aggresive, taking early breaks on the ball, and defenders who are easily followed by misdirection.
In this example, if the quarterback could have immediately thrown the ball to the Y receiver the second he received it, the FS would be measured as having 0.5 seconds in the “Instincts” metric because he switched his coverage to the Y receiver 0.5 seconds (one frame) after the target was determined. Note that the Instincts value can be negative, meaning that a defender was able to anticipate the target before the ball is even thrown.
Cornerback play is very difficult to assess by standard statistics. Interceptions have an element of luck to them and a cornerback with many tackles is often a player who is simply targeted more often, typically a sign of being the weaker link in the defense. The ability to stay in tight coverage, however, is a universal trait of all elite cornerbacks. With that in mind, we have developed the “recovery” metric. It quantifies the average amount of time a player spends in the trailing position before recovering back into either an attack or surveillance motion. Note, although designed with cornerback play in mind, this metric can also be useful in assessing coverage skills of any position and possibly a player who is simply too slow.
The defense plays well on this play, with only the SCB and SS ever falling behind into the trailing motion, both with an average “Recover” time of 0.5 seconds.