Spatio-temporal pattern mining in sports video

Abstract

Sports video is characterized with strict game rules, numerable events and well defined structures. In this paper, we proposed a generic framework for spatio-temporal pattern mining in sports video. Specifically, the periodicities in sports video are identified using unsupervised clustering and data mining method. In this way sports video analysis never needs priori domain knowledge about video genres, producers or predefined models. Therefore, such framework is easier to apply to various sports than supervised learning based approaches. In this framework, a hierarchical spatial pattern clustering routine, including scene-level clustering, field-level clustering and motion pattern clustering from top to bottom, is designed to label each subshot with coherent dominant motion. Then the temporal patterns are identified from such label sequence using data mining method. These mined probabilistic patterns are presented as basic structural elements of sports video. © Springer-Verlag Berlin Heidelberg 2004.