​CEA-List is contributing to a French national research agency (ANR) project called Team Sports to develop video analysis tools to help coaches assess how well their team members play together. Accurate data on players' positions and trajectories at all times during play is vital to generating insights on group dynamics. But video footage of games presents a number of challenges to effective tracking. Players move fast, can be hidden behind other players, and sometimes move outside the camera's field of view. Visually-similar jerseys only add to the complexity.

A university in Belgium asked scientists to come up with a solution capable of tracking players in 30-second video clips as part of the SoccerNet international challenge. Researchers at CEA-List applied neural-network-based video analysis algorithms to the task. First, an algorithm detects the players present in a given frame. A re-identification algorithm then encodes each player's appearance so that individual players can be distinguished from each other in the following frame. Players' appearances and location logic are used to track each player accurately.

The system correctly tracked individual players from one frame to the next 93% of the time. CEA-List won first prize in the SoccerNet challenge, beating out a dozen other contenders from around the globe. The researchers working on the technology will now tackle real-time, 3D position tracking in the field—rather than from video footage—using multiple-camera data fusion. This advance will find applications beyond sports. It could enable new video surveillance solutions, for example.