As we inch closer to the start of NFL Training Camp, we are flooded with stories of players doing what it takes to improve. Getting in better shape, getting stronger, getting leaner, and getting faster are all common talking points across the league. Inevitably, though, the focus turns behind the center to the quarterback position.
Patching mechanical issues, mastering footwork, and speeding up processing are common places quarterbacks hope to improve. Quarterback development is central to NFL franchises, and technological advances can support teams along the way. Leveraging technological advancements is becoming more widespread in quarterback development, but the NFL still lags behind other sports.
In The MVP Machine, written by Ben Lindbergh and Travis Sawchik, the authors discussed how technology was a driving force in baseball’s player development. Driveline Research was helping pitchers enhance velocity through Edgertronic’s high speed video cameras. Analysts were able to collect data on what exactly each joint was doing in a pitching motion and then put in place specific drills to increase pitcher performance while mitigating injuries.
Naturally, my focus turned from this innovative work with pitchers to the stressful act of throwing a football. This technology may not be a magic wand for throwers who are all unique in their own way, but understanding biomechanics and potential pain points may be able to help those who ask so much of their bodies.
There are many challenges in translating Driveline’s research from a baseball diamond to inside the painted lines of a football field. For one, an Edgertronic camera can cost anywhere from $5,000 – $7,500 (here is one listed by Driveline) per camera. In lieu of this option, I started to research the data collection process behind Edgertronic cameras and came across methods and applications related to pose estimation using Computer Vision.
Pose Estimation uses deep learning to predict keypoint locations on the body in a 3D plane (X, Y, Z coordinates). Computer Vision is an area of machine learning that allows you to learn and extract data from digital images and videos. Combining the two would mean being able to extract keypoints, such as joints, from a video or still image and convert it to a data set for analysis.
There are open-source models and methods that have been built to allow you to extract coordinates of keypoints (elbow, knee, shoulder, etc.) such as Carnegie Mellon’s OpenPose project. The data extraction process using pose detection allows you to get coordinates at a given frame and begin to glean insights regarding the location of that joint.
Justin Herbert has solid throwing mechanics, which creates his zip on the ball (~43mph wrist velocity in this throw).
— Quinn MacLean (@QuinnsWisdom) August 22, 2022
One noticeable advantage that baseball has is a single subject is in the foreground of a video to collect data on. I have been collecting data on publicly available throws where the subject in question was clearly separated in the video, but these instances are limited.
To solve for this, I turned to the NFL Combine where you could more easily isolate a thrower and began to collect pose data. There are also other public videos from Pro Days and OTAs that isolate the thrower.
I have been able to collect joint estimates on 110 throws from this criterion. Here is a sample image of how the pose estimation applies to the body. We can see an approximate match of each joint near the point of release.
In this image, focus on the relationship between the throwing elbow and shoulder. In observing the distribution of the elbow angles, I found 3 discrete buckets of the elbow angle: “Ball Held” (<80 degrees), “Windup/Follow Through” (> 110 degrees), and “Release Point” (80-110 degrees). The 3 buckets provide an elbow angle estimate based on all throws in our sample.
The release elbow angle will vary by the quarterback’s arm slot, or relative angle of the arm to the ground, at the release point. Quarterbacks are unique in where their arm slots are as well as the path the ball takes to get to that spot.
Coaching mechanics can be tough because of a person’s muscle memory but also because of the different arm slots a thrower may be comfortable with. A fast-paced game can require different arm slots, and players do not have the luxury of slowly thinking through their mechanics as defenders attack the quarterback.
This gets into the relative position of the throwing elbow in relation to the throwing shoulder. Although we could expect a difference in arm slot from quarterback to quarterback, we generally expect the throwing elbow to be slightly above that of the throwing shoulder.
Below shows the relationship between the elbow and shoulder at approximate release point. We can see a linear relationship with the elbow having a higher height (height here is in pixels and a conversion of x, y coordinates from a 3D plane to a 2D plane) than that of the shoulder.
The Sport Performance U highlighted two main red flags in a throwing motion that can lead to injuries. The first is a low elbow through the release (low elbow = below ear height). The second is having a straight elbow at release. This puts more strain on the shoulder to generate power.
Throwing elbow and shoulder happen to be the spot along the throwing motion where the most quarterback injuries come from. Optimizing for both can reduce the biomechanical strain on the quarterback and can add longevity to a career.
In validating the importance of the shoulder angle, I fit a random forest model on each frame of the 110 throws (~10k data set size) predicting the release point classification. As anticipated, the right shoulder angle or throwing shoulder angle (all players in the data set were right-handed; left hand below would translate to non-throwing hand) was the most predictive of elbow angle.
Another important insight from the model is the importance of having an active lower body when throwing the football. Having good footwork, stance (knee angle, ankle velocity, slope or lean of QB, and ankle distance), and swinging your hips to the target when throwing the ball can help to take strain away from the upper half of the body (this is why you see Dak Prescott swinging his hips in pregame warmups).
This is represented in the SHAP analysis, which highlights the contribution of a variable to the model output where higher values represent a more important variable, of classifying release point below highlighting knee angle, ankle velocity, slope (lean of QB), torso size, and ankle distance.
Another important part of the throw is hip angular velocity or ability to rotate hips at release. Not only does hip rotation help at the point of release, but it can also reduce the strain on the elbow. Using the whole body and swinging your hips through to the target can again help preserve the arm for the long term and create more control in the throwing motion.
You could think of this like a golf swing. Golfers want to make sure their hips rotate through impact not only to generate power but also because torso direction influences the direction of the ball. The same theory of movement can apply to the quarterback throwing the ball.
This is a simple way to look at the throwing motion by using pose estimation, but this analysis does come with drawbacks. The main one is that the entire sample size is from one angle per throw. I would recommend getting multiple angles for validation, but in the absence of an implementation of multiple cameras and pulling from public sources, the information described above is what I was able to discover.
This could help in instances where the pose estimation was not able to get a proper reading for a given frame. This is a common issue in marker-less systems which require multiple cameras to capture all angles (normalizing data from a 3D to a 2D scale can help in validation of a marker-less system).
I am hopeful technological investments like we have seen in baseball can make this type of data more available in the future to spark a new era of player development in football and help preserve athletes for longer careers.