Verified by 
Key points
- A new study demonstrates how AI deep learning and computer vision can predict the formation of the offensive team with a high degree of accuracy.
- Researchers presented a method of automatically locating and labeling players and identifying the offensive formation before the play begins.
- The researchers reported that they obtained greater than 90% accuracy on player detection and labeling and 84.8% on identifying formation.
As this Sunday’s National Football League (NFL) Super Bowl LVII approaches, American football fans eagerly anticipate the epic showdown between the Philadelphia Eagles and the Kansas City Chiefs. The stakes are high for the coaches and players who have undoubtedly been analyzing their opponent’s video footage throughout the season.
A new study published this month demonstrates how artificial intelligence (AI) deep learning and computer vision can predict the formation of the offensive team with a high degree of accuracy–a potential game-changer for pre-game preparation for professional and collegiate football and other team-based sports.
“Analyzing and annotating sports footage manually can be tedious and time-consuming, so automating this process has great potential in reducing human error, saving time, and decreasing costs,” wrote Dah-Jye Lee, MBA, Ph.D., professor, and the director of the Robotic Vision Laboratory in the Electrical and Computer Engineering Department at Brigham Young University (BYU), along with BYU researchers Jacob Newman, Andrew Sumsion, and Shad Torrie. “This automated analysis has the potential to assist both coaches and players in understanding how their team plays and how other teams play, which can improve overall team performance.”
This proof-of-concept algorithm developed by machine learning researchers shows the potential of AI serving as a strategy analysis system that may disrupt not only American football but also other team-based sports, such as baseball and soccer, at both the collegiate and professional levels.
“We present a method of automatically locating and labeling players, as well as identifying the offensive formation, from an overhead image of a football play before the play begins,” the researchers wrote.
The AI system design consists of three module systems: player localization, player labeling, and formation identification.
The player localization module finds the location of the visible players in an overhead picture of a football formation. The researchers use still images for this study with a You Only Look Once (YOLO) deep learning framework, which enables video analysis for future studies. Namely, the team used TrainYourOwnYOLO by Anton Meuhlemann from the GitHub repository to train a YOLOv3 architecture with a special custom dataset of images of individual players.
Once the player localization module identifies the locations of the visible players, this data is used as input for the player labeling module, separating the defensive from the offensive players.
The player labeling module is an AI residual network (ResNet) framework commonly used in AI computer vision. Its AI deep neural network architecture has a depth of 152 layers. Twelve player positions were grouped into eight labels consisting of the Offensive Line (Center, Offensive Guard, Offensive Tackle), Quarterback, Running Back, Tight End, Wide Receiver, Defensive Back (Cornerback, Safety), Defensive Line (Defensive End, Defensive Tackle), and Linebacker. The player location and label data are used as input for the formation identification module.
“As college football teams are most interested in analyzing offensive formations, we focused the formation identification module solely on identifying offensive formations,” the researchers wrote.
The formation identification module is also a ResNet with 152 layers in its deep neural network. This module was trained with 25 offensive formations grouped into five families. The five families of formations consisting of the I formation (I Form Close Slot, I Form H Pro, I Form H Slot Open, I Form H Tight, I Form H Wing), Pistol (Pistol Bunch TE, Pistol Full Panther, Pistol Spread, Pistol Strong Slot Open, Pistol Wing Flex), Shotgun (Shotgun Ace, Shotgun Doubles, Shotgun Eagle Trey, Shotgun Slot Offset, Shotgun Wing Tight), Singleback (Singleback Ace Double Wing, Singleback Deuce, Singleback Doubles North, Singleback Trio, Singleback Wing Pair), and Strong (Strong H Pro, Strong H Slot, Strong H Wing, Strong Tight, Strong Twins Over).
“We obtain greater than 90 percent accuracy on both player detection and labeling, and 84.8 percent accuracy on formation identification,” the researchers reported. “These results prove the feasibility of building a complete American football strategy analysis system using artificial intelligence.”
Copyright © 2023 Cami Rosso All rights reserved.
