Goalkeepers struggle to guess which direction the penalty taker will shoot
Javier Soriano/AFP via Getty Images
Trained with over 1000 penalty kicks in football games, the deep learning model can predict how the ball will work better than the actual goalkeeper.
“Penalty kicks are some of the most decisive moments in football and often determine the outcome of major tournaments,” says David Freire O'Regon at Las Palmas de Gran Canaria University in Spain. “Nevertheless, real-time support for goalkeepers is still based on intuition. We wanted to investigate whether machine learning could predict the direction of a shot from the kicker's body movements.”
So Freire-Obregón and his colleagues cut 1010 penalty kicks from Real, Teleaved Matches in Spain. Of these clips, 640 were considered analyzable by the AI model, with the rest being discarded for being blurred, too short or obstructed.
We then fed each clip to 22 deep learning models. This required us to guess whether the penalty would be left, right or center left, lower right, based on the video footage and the simple fact that the player was right or left foot.
The best performance model was able to correctly identify whether the ball went right, left, or 52% of the time in the middle. When researchers removed the unused middle option, the model's accuracy increased to 64%.
Researchers were surprised, “we can reveal the intent of how subtle movement clues and how the ball is before it's kicked,” says Freire-Obregón. He hopes that this information will be useful for training goalkeepers, but he hopes that using AI predictions will be more difficult in match situations.
“What we're looking to explore next is whether these odds can be predicted in advance for a penalty kick, using only the kicker's movement before the shot,” he says. “If so, how quickly can you make such predictions while maintaining meaningful accuracy?”
topic:
