Description »

As part of our people tracking from multiple cameras research, we track and analyze the behavior of basketball players. The detection and tracking are based on synchronized and calibrated cameras. The current tracking algorithm does not take into account the “motion model” of the multiple players.

Sources of uncertainty in tracking include ambiguous image measurements and inaccurate motion models. In the recent years, motion models have been proposed that take into account social forces, e.g., the desire to stay with a group. Empirical evidence suggests that these models improve tracking accuracy in crowded scenarios. We would like to investigate motion models suitable for sports.

The first objective of this project is to investigate the shortcomings of the existing models. For example, current social models are hand-crafted based on common-sense knowledge of pedestrian interactions. Therefore, these models are not likely to be useful for tracking players in team sports, e.g., basketball players, where players deliberately block each others' path.

The second objective is to propose improvements to the existing multi-person dynamical models. For example, is it possible to automatically discover short-term (five-seconds) interaction patterns for 2, 3, or 4 players?

Given such patterns, is it possible to automatically learn a composition of such patterns for a team sport?

Can such model be used to improve tracking and to alert the viewers to an unusual (creative) play?

If successful, the developed application will be used within an industrial project which involves international basketball championships.