Sports Highlights Generation using Decomposed Audio Information
Muhammad Rafiqul Islam, Manoranjan Paul, Michael Antolovich, Muhammad Ashad Kabir · 2019
Visual features are generally used in video summarization including sports videos. However, audio features are also important in summarization as they characterize the exciting events of a sport, especially when external thing e.g. raining effects the visual feature. The commentator's voice, referee's whistle, and audience-voice are strongly correlated to the exciting events e.g. goal, penalty shootout, red card, etc. in a soccer game. The existing methods using original audio signals could not provide expected summary due to the various noises from the audience such as horn and sound from various musical instruments. In this paper, we propose a novel approach to extract audio features using Empirical Mode Decomposition (EMD) that can filter out noises. We have conducted a set of experiments on four real soccer videos. The experimental results demonstrate that the proposed method can detect goal events with 96% accuracy which outperforms the existing state-of-the-art method by 42%.