Automated full-game ice hockey analytics

点击翻译按钮获取中文摘要

Computer vision is a research field that involves creating algorithms capable of interpreting scenes. A key challenge is the automatic generation of analytics from digital video to mimic the ability of humans. Although pop culture gives the impression that engineers have created such algorithms to near perfection, the reality is that we are still challenged to create solutions that are efficient, effective, and consistent even for dedicated applications. Generating analytics from broadcast video of team sports is one such application. For ice hockey, human captured analytics typically focus only on puck-centric events (e.g., goals, shots) and it is not feasible for humans to efficiently interpret all game activities and events. To capture comprehensive analytics, computer vision can be developed to interpret all player activities and then use this information to interpret game events. Such information can be used to process game videos far more efficiently than a human operator in support of scouting opponents, preparing team game plans, assessing draft selections, and preparing players for games. The challenges with using computer vision to capture hockey analytics involve the speed of the game, the inability for broadcast video to identify the small puck, the persistent occlusions of players/pucks (by the boards and other players), identifying unique players that have similar builds and same uniforms, and limited field-of-view of the video. Our research team has made strong progress in methods to learn locations of players on the ice, tools for capturing ground truth data, tracking all player movements simultaneously, and working in partnership with a company with a vested business interest that already provides data to professional leagues. Our objective is to advance the state-of-the-art in hockey analytics to be able to automatically interpret all game events and team strategy for a single broadcast ice hockey game video.