Researchers at the University of Waterloo’s Vision and Image Processing (VIP) Lab have developed a way to analyze videos of hockey games using artificial intelligence (AI).
The team responsible for the development includes Kanav Vats, a PhD student in systems design engineering, along with his doctoral supervisors, professors David Clausi and John Zelek, as well as postdoctoral fellow Mehrnaz Fani.
Their AI system, which combines two deep-learning AI techniques, can identify hockey players by their sweater numbers with 90 per cent accuracy.
Currently, sports commentators can struggle to quickly identify sports players because footage may be taken from a distance to show the broader progression of the game, and the speed of players and the cameras following them can increase motion blur.
According to Vats, “the only major cue you have to identify a particular player in a hockey video is jersey number. Players on a team otherwise appear very similar because of their helmets and uniforms.”
Members of the VIP Lab are working on player identification as well as other projects to analyze player performance in partnership with Stathletes Inc., a company that provides industry-leading data and analytics about sports games and athletic performance. Their custom software tools track and record every aspect of hockey games.
The VIP Lab, which is housed under the systems design engineering department, supports research related to visual processes. The lab is dedicated to “finding solutions for the outstanding problems in visual perception and processing,” according to its website.
Since its formation in 1980, the VIP Lab has supported research in a range of topics including vision models, perception, image and signal processing, pattern recognition, machine learning and AI. Information about their current research topics and research demos is available online.
For the hockey project, the research team trained AI algorithms to recognize sweater numbers in new images using a set of more than 54,000 images — the largest data set of its kind — from National Hockey League (NHL) games.
They used an approach called multi-task learning, through which accuracy was boosted by representing multiple-digit numbers such as 12 as both a two-digit number and as two single digits — one and two — put together. With multi-task learning, machines attempt to learn multiple tasks simultaneously, rather than independently learning each task.
The team’s research will likely simplify the responsibilities of people working to annotate hockey games, which is currently an arduous and time-consuming task. “As you can imagine, a person manually annotating a video of a full hockey game of three periods would take hours,” Vats said. “Machine-learning systems can produce data from videos in a matter of minutes.”
In the future, the researchers said they expect their technology could be transferred to other team sports, with modifications.