AbstractsComputer Science

Extracting ground plane location data of rugby 7s players, from a zooming and rotating camera

by J van Dis

Institution: Universiteit Utrecht
Year: 2016
Keywords: computer vision; ground plane multi target tracking; camera modelling; camera pose estimation; Kalman filter; deformable parts model; object-detection
Posted: 02/05/2017
Record ID: 2118440
Full text PDF: http://dspace.library.uu.nl:8080/handle/1874/328396


Finding the trajectory of sports persons in videos can be done in various ways. We address this problem for rugby 7s under difficult conditions, such as occlusion, a small camera angle with respect to the horizon, zooming, and supporters standing nearly in the field. A methodology is presented that processes videos of rugby 7s games during which simultaneously the camera parameters are derived - stating its location, orientation and internal settings such as focal length, and the position of the rugby players within the image plane are detected, to tracks their movements with respect to the ground plane. We propose a new method to determine the camera parameters effectively, making use of the recurring nature of the camera state, by only using a limited set of reference frames, called baseframes. Tracking is performed in the ground plane, using a tracking-by-detection approach with a Kalman filter to dynamically model the tracklets. An evaluation of our methodology shows the potential of our methodology. However, it is recommend to improve of the conditions under which tracking is performed. Advisors/Committee Members: Poppe, R.W., Dignum, F.P.M..