Researchers at Disney have developed new robotic cameras that mimic human operators to anticipate action during a basketball game for better frame shots.
Automated cameras make it possible to broadcast even minor events, but the result often looks robotic.
Now scientists have made it possible for robotic cameras to learn from human operators how to better frame shots of a basketball game.
Many automated systems determine where to point the camera by tracking a key object.
But human camera operators are able to anticipate action and can adjust the camera's pan, tilt and zoom controls to allow more space, or "lead room," in the direction that the action is moving.
The result is video imagery that is smooth and aesthetically pleasing, researchers said.
Peter Carr, a Disney Research engineer and Jianhui Chen, an intern and a PhD student in computer science at the University of British Columbia, devised a data-driven approach that allows a camera system to monitor an expert camera operator during a basketball game.
The automated system uses machine learning algorithms to recognise the relationship between player locations and corresponding camera configurations.
"We don't use any direct information about the ball's location because tracking the ball with a single camera is difficult," Carr said.
"But players are coached to be in the right place at the right time, so their formations usually give strong clues about the ball's location," Carr added.
Carr and Chen demonstrated their system on a high school basketball game. They used two cameras - a broadcast camera that was operated by a human expert and another that was a stationary camera that the computer used to detect and track the players automatically.
"Because the main broadcast camera in basketball maintains a wide shot of the court, we focused on predicting the appropriate pan angle of the camera," Carr said.
Following supervised learning based on the operator's actions, the system was able to predict how to pan the camera in a way that was superior to the best previous algorithm and that did indeed mimic a human operator.
Carr said he expects the method can be adapted to other sports, possibly with additional features.
