Feature design in soccer video indexing

Unlike American football, baseball, tennis and many other sports games, soccer is not a well-structured game. Soccer videos are basically continuous streams with exciting highlights embedded. However, the highlights of the same type are correlated in spatial and temporal feature distributions. In this paper we present an effective scheme to represent soccer scenes with low/mid-level image and sound features. We discuss three aspects of feature design in soccer video indexing system: temporal structure, low/mid-level features, domain specific knowledge. We use the maximum-entropy based machine learning method as a test platform to verify the feature design scheme. The maximum-entropy method can automatically choose the features with more distinguishing power. The feature representation is applied to soccer video indexing. Extensive experiments are conducted and satisfying results are described.