Semantic event detection using hidden markov model in golf video

Abstract

As the amount of multimedia has been rapidly increased with wide a spread of Internet service and multimedia technologies, users are likely to consume multimedia contents efficiently and actively. Video content analysis to provide various services such as video indexing, retrieving and abstracting has become important. In the video content analysis, semantic event detection is needed for a practical application. But most of semantic event detection algorithms are dependent on video contents and required heuristic parameter determination. In this paper, we propose an event detection algorithm using Hidden Markov Model (HMM) in golf video. HMM is suitable to characterize the temporal pattern of semantic events as well as the instantaneous characteristics of each event because it is a probabilistic model based on Markov process. To apply an HMH in golf video, we defined three semantic events: ``Drive``, ``Bunker``, and ``Putting``. Then we modeled the HMMs for each semantic event. Finally, with the models of each event, we adopted MPEG-7 visual features to distinguish each state in HMM. Compared with previous semantic event detection, the proposed algorithm has two advantages. One is that the proposed method could provide a generalized event detection method which is robust for various contents. The other is a high accuracy of detection ratio to determine the segment boundary of detected events. To prove the usefulness of the proposed algorithm, we performed experiments with four golf videos; each of the videos is about 40 minutes long. The experiment results show that the accuracy of the event detection is about 80% on average. Also we could observe that segment boundary of detected events are more accurate than the previous methods.