While the best soccer players on the planet take to the pitch in this summer’s World Cup, amateur coaches can reach for the stars by adding AI to their playbooks.
Veo, a Copenhagen-based startup, has created a camera for its users to place midway on a soccer field touchline (that’s the “sideline” for U.S. readers) that captures the action across the entire field. The company is using its own object-detection model to track movement of the soccer ball to follow the game play.
Founder and CEO Henrik Teisbaek said Veo was built to offer professional video analytics capabilities widely. While the soccer elite play in Russia for the quadrennial soccer competition, Veo has the globe’s 100,000-plus amateur clubs battling year-round in its sights.
“These coaches are literally on chalk boards from the 1950s. Veo is for the coaches that aspire to be professional,” Teisbaek said.
Veo’s cool-looking, lime-green triangular device packs two 4K video cameras inside for use on a tripod. The cameras require their data to be transferred to Veo’s cloud service, which stitches footage together to produce panoramic videos that track the game action.
Coaches can use the processed videos to zoom in on the ball and players for post-game analysis. The system also allows coaches to place lines on the field and other marks to help analyze in the videos whether players are in formation with their practice drills.
Started in 2016, Veo is now working with 20 soccer clubs in Copenhagen. The first generation cameras that it has operating on the soccer fields capture about 100 matches a month.
“That is about four or five times what is broadcast on TV in Denmark,” Teisbaek said.
Veo is now developing its next-generation camera, which will launch later this year. Close to 100 clubs have already bought the cameras in presales. The cameras sell for about $1,200 Euros.
Behind the scenes, Veo’s convolutional neural network acts as a big brain for video processing. The neural network has been trained on 1.5 million images consisting of roughly 1 million soccer balls and 500,000 players.
Veo’s heavily trained image model makes it possible to quickly find game highlights, such as a scored goal. But the company is continuing to train its model to offer more capabilities.
Veo processes about 500 million pixels per second. “We use NVIDIA processors to accelerate the process. That’s what’s needed in order for us to go to real-time processed videos,” said Teisbaek.
In addition to enlisting GPUs from a German data center, Veo uses the NVIDIA CUDA Deep Neural Network library (cuDNN) with TensorFlow.
Veo isn’t alone in AI-driven sports analytics. Toronto-based startup ICEBERG is using GPU-powered analytics for Hockey.