Extract Soccer Video Highlights with Modified CRF

Abstract

Soccer video analysis is attracting much attention. In this paper, we propose a modified conditional random field (CRF) model to extract the highlights of an entire soccer video. Highlight extraction in soccer video is essentially a kind of timing annotation problems, so the commonly used CRF model is adopted to solve this problem in this paper. Meanwhile, Boolean function is normally used as feature function in the CRF model, which will result in a hard association between observed variables and highlight variables. By introducing Bayesian network to model the observed variables and replacing the original feature function with posterior probability calculated with Bayesian network, hard association is transformed into soft association, which makes the model more close to the actual situation. Experimental results show that the proposed algorithm has achieved good results.