Soccer Event Retrieval Based on Speech Content: A Vietnamese Case Study

Abstract

Video is a self-contained material which carries a large amount of rich information, far richer than text, audio or image. Researches (Amir et al., 2004), (Fleischman & Roy, 2008), (Fujii et al., 2006) have been conducted in the field of video retrieval amongst which contentbased retrieval of video events is an emerging research topic. Figure 1 illustrates an ideal content-based video retrieval system which combines spoken words and imagery. Such ideal system would allow retrieval of relevant clips, scenes, and events based on queries which could include textual description, image, audio and/or video samples. Therefore, it involves automatic transcription of speech, multi-modal video and audio indexing, automatic learning of semantic concepts and their representation, advanced query interpretation and matching algorithms, imposing many new challenges to research. There is no universal definition of video event, and the existing definitions can be classified into two types: one is being abnormal and the other is interesting to users (Babaguchi et al., 2002). In the first type of definition, an event may be either normal or abnormal. Generally speaking, only the abnormal event, which has more information than the normal one, is meaningful to the users. This event definition is suitable for the video analysis under restricted circumstance such as surveillance. The event definition of interesting to users is based on the users’ description and domain prior knowledge (Sun & Yang, 2007). Suitable examples of this category are sport-video events such as ones in soccer and baseball. Several popular soccer events are shown in Fingure 2, including scoring, corner kick, yellow card and foul events. Soccer video analysis plays an important role in both research and commerce. The basic idea of soccer events retrieval is to infer and retrieve the interesting events, and its goal is to make the results accord with human’s visual perception as much as possible (Xu et al., 2001). Inference of events can be stemmed from either the semantic visual concepts or the spontaneous speech embedded in the videos. This chapter approaches soccer-video event retrieval in an audio aspect (i.e., the problem of spontaneous speech recognition). In this case, an event is defined as the spatiotemporal entity interesting to users, which is remarked by the announcer’s spoken words. By exploiting spoken information of the video, soccer events are detected using an automatic speech recognition (ASR) system. However, as soccer videos vary in both speech quality and content, a canonical speech recognizer would not perform well without modifications and improvements. There are three main problems