Semantic event detection in baseball videos based on a multi-output hidden Markov model

Abstract

In this paper, we proposed an event detection method in baseball videos based on a multi-output HMM (hidden Markov model), using high-level audio/video features. For the video part, we use eight kinds of semantic scenes detected from baseball videos in our previous work. For the audio part, we extract the audio shots from corresponding video scenes, and cut an audio shot into N one-second clips. Then, the MFCC and ZCR of a one-second clip are extracted and fed into the SVM for classifying it as acclaim and silence. Based on the classification results, the type of an audio shot can be determined in the post-classification. Next, a multi-output HMM modified from the original HMM is used to combine video and audio features to detect baseball video events. Finally, the experimental results show, the multi-output HMM has good event detection accuracy.