Abstract:
Sports related analytics have become a main component of the present professional
sporting domain. Teams continuously rely on the knowledge provided by
analytics systems to gain a competitive edge over the opposing team. One of
the main aspects of sports analytics is automated player tracking which can be
achieved by computer vision based techniques by analyzing video footage of sporting
events. Multiple object tracking in itself is a non trivial problem due to the
large number of variables involved. This is further amplified by the high number
of occlusions, trajectory changes that occur in a highly physical sport such as
Rugby. We set out to solve the problem of automated player tracking using a
tracking by detection approach. We make use of an object localisation model
named YOLO and retrain it to suit the specific scenarios in Rugby. In order to
solve the data association problem we compute an appearance based metric using
an identity embedding encoder network. A Kalman filter is used along with the
appearance based metric to establish the associations between tracks and detections.
We conduct several experiments to evaluate the implemented solution and
report the results. We discuss the limitations,further improvements and areas
that present further research opportunities.
