Claasen Paul J. ; de Villiers J. P.

Abstract

A novel Bayesian framework is proposed, which explicitly relates thehomography of one video frame to the next through an affine transformationwhile explicitly modelling keypoint uncertainty. The literature has previouslyused differential homography between subsequent frames, but not in a Bayesiansetting. In cases where Bayesian methods have been applied, camera motion isnot adequately modelled, and keypoints are treated as deterministic. Theproposed method, Bayesian Homography Inference from Tracked Keypoints (BHITK),employs a two-stage Kalman filter and significantly improves existing methods.Existing keypoint detection methods may be easily augmented with BHITK. Itenables less sophisticated and less computationally expensive methods tooutperform the state-of-the-art approaches in most homography evaluationmetrics. Furthermore, the homography annotations of the WorldCup andTS-WorldCup datasets have been refined using a custom homography annotationtool that has been released for public use. The refined datasets areconsolidated and released as the consolidated and refined WorldCup (CARWC)dataset.