| Summary: | The goal is to automatically determine puck possession and location in broadcast hockey video. The task is challenging for several reasons. The puck itself is small, moves rapidly and lacks distinctive local visual features. Players also move and change direction rapidly. Interactions among players on the same team and with players on the opposing team are complex. This thesis demonstrates an automatic system composed of two modules that explore the task from different perspectives. The system takes as input player tracking results that have been registered to a geometric model of the ice surface under a given homography. It analyzes video sequences according to the user specified module and generates graphical annotation and data storage as output. In the first module, puck candidates are detected directly. An innovative hierachical graph-based method is used to track puck transitions between players. Based on these puck transitions, a puck state engine is created to determine puck possession and to select the most likely puck trajectory. In the second module, a dense player motion field is constructed to estimate puck location and possession according to motion convergence points. This work is a first attempt to extend automated hockey video analysis to include puck detection and puck possession. It stands as a proof of concept system for the broad area of team sports analysis.
|
|---|
