A generic approach to classify sports video shots and its application in event detection

Abstract

Shot type is useful information for semantic sports video analysis. Most existing approaches utilize predefined rules and domain knowledge to derive shot types in sports video. Although these methods have achieved promising results in some specific games, it is hard to extend them from one sport to another. To address this problem, we propose a generic approach to classify shots in sports video. Our approach utilizes bag of visual words model to represent key frame for each shot based on Scale Invariant Feature Transform (SIFT) feature points; either Support Vector Machine (SVM) or Probabilistic Latent Semantic Analysis (PLSA) are then employed to classify key frame to determine shot type. As our approach relies little on domain knowledge, it can be more easily extended to different sports. We have evaluated our shot classification approach over five types of sports video and have achieved promising results. To show the usefulness and effectiveness of our shot classification, we apply the results of shot type to detect events in basketball video via a generative-discriminative model. In addition, we have observed that some common visual parts frequently appear across various shots in the same sport or even different but relevant sports. For instance, soccer and basketball are relevant sports in the sense of field-ball game. Motivated by this observation, we attempt to alleviate the problem of insufficient sports video data in some applications by sharing these visual parts across different but relevant kinds of sports.