How AI can track hockey action from faceoff to finish

Researchers at the University of Waterloo have developed two innovative artificial intelligence (AI) systems that significantly improve how hockey games can be analyzed using video footage without the need for expensive equipment.

Drawing on Waterloo's strengths in computer vision and systems design, the research advances the evolving field of automated sports analytics. The new tools address long-standing challenges in tracking fast-moving game action, including obstructed views and motion blur common in broadcast feeds.

"These improvements in detection accuracy could transform how coaches, teams, and broadcasters analyze game dynamics, leading to better strategic decisions and more engaging fan experiences," said Dr. David Clausi, a professor of systems design engineering at Waterloo.

Engineering professors Dr. David Clausi (left) and Dr. John Zelek play pickup hockey together and lead research at the University of Waterloo into how to better track and analyze the game using AI tools. (University of Waterloo)

In one study, researchers developed a model that leverages the fact that players typically keep their eyes on the puck during games to help infer its location based on body position and the direction of their gaze.

The AI-based system, called Puck Localization Using Contextual Cues (PLUCC), boosted the accuracy of puck location by 12 per cent, while also reducing localization error by more than 25 per cent compared to existing technology.

Researchers expect the system to be particularly useful for smaller organizations and amateur teams by offering a low-cost alternative to much more elaborate and expensive tracking technology such as Hawk-Eye.

"Our goal was to make puck tracking something that doesn't require a million-dollar setup," said Liam Salass, a graduate student who was lead author of the study. "If a coach can analyze a game using only video, that's a big win for accessibility in sports analytics.

"Finding the puck in broadcast video is one of the toughest problems in sports vision, so seeing our system accurately predict its location using contextual cues was incredibly rewarding. It was like we'd given computers real game sense."

The second study involved the development of an AI-based framework called SportMamba that improves how multiple moving players are tracked in sports videos. The model dynamically predicts player movements during games, accounting for rapid motion, blocked camera angles and camera shifts.

Tested across soccer, basketball, and hockey footage, SportMamba outperformed existing tracking methods by up to 18 per cent in accuracy and efficiency, allowing teams and broadcasters to conduct real-time, data-driven performance analysis without the need for costly sensor systems or fixed-camera setups.

"Tracking a hockey player on a breakaway is relatively easy," said Dr. John Zelek, also a systems design engineering professor and a director with Clausi of the Vision and Image Processing (VIP) Lab at Waterloo.

"It's much more difficult to track and differentiate players in a scrum along the boards or in front of the net. SportMamba can tackle these difficult situations and tell us, for example, who deflected the puck and scored."

The research papers, "Ice Hockey Puck Localization Using Contextual Cues" and "SportMamba: Adaptive Non-Linear Multi-Object Tracking with State Space Models for Team Sports," were recently presented during the 2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops.

Featured image: University of Waterloo Engineering master's student Liam Salass developed an AI-based system that improved puck detection when analyzing game video. (University of Waterloo)