System leverages machine learning to overcome challenges of speed and visibility in fast-paced hockey games.
On Nov. 6, the University of Waterloo announced that its researchers had developed a new artificial intelligence system capable of tracking a hockey puck using nothing more than regular game footage and contextual cues from the players on the ice.
The system, called Puck Localisation Using Contextual Cues (PLUCC), was created by master’s student Liam Salass under the supervision of professors David Clausi and John Zelek from the University of Waterloo’s Department of Systems Design Engineering. Unlike the expensive sensor- and multi-camera-based tracking technology used in professional leagues, PLUCC uses a “detection network” paired with a “contextual encoder” that predicts the puck’s location based on players’ gaze, body position, stick placement, and other visual cues in broadcast video.
“The contextual encoder essentially learns where the puck might be based on what the players are looking at, their pose, and their stick position,” Salass explains.
Tracking the puck in live hockey broadcasts has long been a difficult problem due to its small size, high speed, and frequent obstruction behind players. Traditional systems like the NHL’s sensor-based tracking require costly infrastructure, making them inaccessible to most leagues. PLUCC aims to offer a low-cost alternative that could unlock advanced analytics for junior teams, university programs, lower-tier leagues, and even community hockey.
According to Salass, PLUCC already outperforms many “off-the-shelf” small-object detection systems and can reliably detect puck location, offering insights such as puck speed, passing sequences before goals, and heat maps showing where the puck spent the most time.
“I think a team can possibly be at an advantage if they’re using this tool correctly to review their videos,” he says.
Rather than track players — a more complex problem involving the identification of athletes in similar uniforms — Salass focused on the puck, believing contextual cues would carry stronger predictive power. Early experiments confirmed that when trained only on segmentation masks of players, excluding the rink and puck entirely, the model still learned to infer where the puck likely was based on collective gaze direction. That insight shaped PLUCC’s architecture as a hybrid of direct detection and contextual reasoning.
In its current form, PLUCC analyses individual video frames, which means it can miss the puck when visual or contextual evidence is weak. Salass is now developing a version that processes full video sequences, enabling continuous inference even when the puck momentarily disappears behind players.
Beyond hockey, Salass sees the system being used in other sports such as lacrosse, where small, fast-moving objects are similarly difficult to track. He also imagines future integrations with multi-camera systems, long-form video analytics, and AI tools that could one day reconstruct plays in 3D or allow “what-if” simulation scenarios to play out.
Experts watching the rapid evolution of AI in sport are cautiously optimistic but emphasize that the technology is still in its early stages.
“I think it is too early to predict the impact of AI on sport performance,” said Brian Bourque, the head coach of the men’s hockey team at the University of Waterloo and Associate Athletics Director. “With that being said, tools are becoming available which allow AI to ‘cut’ game video quite quickly, which allows coaches to access critical footage immediately after games. It will be interesting to see the development of AI and its role in the future of sports.”
One potential application of PLUCC is in scouting and training, where new analytics — such as “net visibility” before a shot — can help coaches assess shot quality, goalie positioning, and decision-making. These forms of analysis were previously extremely time-consuming or impossible for most teams without expensive tracking equipment. By using standard broadcast footage, PLUCC can generate advanced breakdowns within minutes or hours rather than the days or weeks required for manual review.
For professional leagues, existing sensor networks may remain the gold standard, but PLUCC’s significance lies in democratizing high-level analytics. For the thousands of teams that cannot afford specialized hardware, the University of Waterloo’s work demonstrates how AI could make advanced performance analysis accessible, affordable, and immediate — pushing sports closer to a data-driven future.
Contributed Photo/Sangjun Han
