Unsupervised clustering of dominant scenes in sports video

摘要

We propose a new and efficient approach for clustering dominant scenes in sports video. To perform clustering in an unsupervised manner, we devise a recursive peer-group filtering (PGF) scheme to identify prototypical shots for each dominant scene, and examine time coverage of these prototypical shots to decide the number of dominant scenes for each sports video under analysis. To improve clustering efficiency, we employ principal component analysis and linear discriminant analysis to project high dimensional shot features to lower dimensional spaces suitable for classification. The main contribution of the paper lies in the formulation of clustering dominant scenes in sports video and the development of an efficient, unsupervised solution making use of PGF, time-coverage criterion, and subspace analysis. Experimental results obtained from various types of sports videos are presented to show the efficacy of the proposed approach.