ETH Research Could Democratize AI in Football

(CONNECT) In the future, artificial intelligence in football should not only be reserved for major football tournaments that can afford an expensive computer-assisted multi-camera system. To this end, the Adavanced Interactive Technologies Lab at ETH Zurich has developed a system for Zurich-based FIFA that, according to a press release, requires just one camera instead of ten to twelve.

In 2021, football's world governing body wanted a data set that could be used to train computers to estimate poses. He then provided ETH with video sequences of the 2022 FIFA World Cup in Qatar, which had been recorded with various cameras. It took the ETH team a total of three years, much longer than expected, to present its WorldPose data set at the European Conference on Computer Vision in Milan. The technical challenges included players obscuring each other, motion blur, the zoom and problems with the calibration of the cameras.

"We wanted to find an algorithm that is also accurate enough for greater distances," says computer science doctoral student Tianjian Jiang. WorldPose now contains over 2.5 million individual player positions in 3D. This makes it possible to record all outfield players from both teams at the same time and analyze where they are and what they are doing with or without the ball. This data is now fed into a mobile transmission camera, which delivers around 70 percent of all transmitted images in each stadium.

Since, according to Jiang, "a lot of research is still needed for a functioning and stable system", FIFA has now launched an Innovation Challenge. She will continue her research with the ETH WorldPose dataset. 150 participants from all over the world have already registered. "We will also continue to work on the data set and develop further models ourselves," says Jiang. ce/mm