Tracking and modelling of team game interactions

Team games are complicated activities, which involve much interaction between players. Analysing these interactions is of considerable interest; however it is time consuming, error prone, and unreliable to manuallly obtain the positions of the players throughout a game. Automating this process could produce efficient and repeatable results.

Multiple object tracking in large, congested, rapidly changing, and frequently occluded domains (for example a soccer pitch) is a complex problem, particularly given the non-linear nature of each player's movements. This thesis presents a stochastic sampling based multiple object tracker, capable of tracking objects from a single camera, in the complex domain of sports games. Sports players' shapes vary dramatically, presenting challenges to existing techniques. Multi-resolution template based feature descriptors are learned from example players' shape data, providing a mechanism for identifying players' locations in images. Sports scenes are often busy, in the sense that there may be many players close to each other, causing occlusion of one or more players. the use of multiple cameras to resolve these ambiguities is investigated.

Performance evaluation of computer vision systems is an important and often understudied activity. The performance evaluation in this thesis focuses on postional performance evaluation. New metrics and statistisc are presented which provide an important insight into how well a tracking system is performing (and why it may not be).

Analysing the movements of players over time allows a behaviour model of their movements and interactions to be learned. Positional player data is represented and described using density estimation methods. An emergent appraoch to identifying players is presented. Each player's response from a set of learned Gaussian mixture models is used in a graph partitioning scheme. Thsi allows the identification of each player's 'position'. Behaviour and interaction models have potential uses for analysing tactics, identifying good or atypical players, and most powerfully to be incorporated into a multiple obkect tracking system to govern the expected dynamics of the players.