Towards Efficient Ice Surface Localization From Hockey Broadcast Video

Abstract

Using computer vision-based technology in ice hockey has recently been embraced as it allows for the automatic collection of analytics. This data would be too expensive and time-consuming to otherwise collect manually. The insights gained from these analytics allow for a more in-depth understanding of the game, which can influence coaching and management decisions. A fundamental component of automatically deriving analytics from hockey broadcast video is ice rink localization. In broadcast video of hockey games, the camera pans, tilts, and zooms to follow the play. To compensate for this motion and get the absolute locations of the players and puck on the ice, an ice rink localization pipeline must find the perspective transform that maps each frame to an overhead view of the rink.

The lack of publicly available datasets makes it difficult to perform research into ice rink localization. A novel annotation tool and dataset are presented, which includes 7,721 frames from National Hockey League game broadcasts.

Since ice rink localization is a component of a full hockey analytics pipeline, it is important that these methods be as efficient as possible to reduce the run time. Small neural networks that reduce inference time while maintaining high accuracy can be used as an intermediate step to perform ice rink localization by segmenting the lines from the playing surface.

Ice rink localization methods tend to infer the camera calibration of each frame in a broadcast sequence individually. This results in perturbations in the output of the pipeline, as there is no consideration of the camera calibrations of the frames before and after in the sequence. One way to reduce the noise in the output is to add a post-processing step after the ice has been localized to smooth the camera parameters and closely simulate the camera’s motion. Several methods for extracting the pan, tilt, and zoom from the perspective transform matrix are explored. The camera parameters obtained from the inferred perspective transform can be smoothed to give a visually coherent video output. Deep neural networks have allowed for the development of architectures that can perform several tasks at once. A basis for networks that can regress the ice rink localization parameters and simultaneously smooth them is presented.

This research provides several approaches for improving ice rink localization methods. Specifically, the analytics pipelines can become faster and provide better results visually. This can allow for improved insight into hockey games, which can increase the performance of the hockey team with reduced cost.