出版社：The Institute of Image Information and Television Engineers
摘要：We developed a technique for video data mining to extract ‘semantic patterns’ associated with semantically relevant events in videos. First, several types of raw-level metadata are derived from the raw video data in each shot. The metadata is then sequentially aggregated into a multistream. Then, sequential patterns, each of which is a temporally ordered set of raw-level metadata, are extracted from the multistream. The sequential patterns are reduced to likely semantic patterns using two types of time constraints that we introduce: ‘semantic event boundaries’ and ‘temporal localities’. We developed a prototype system and demonstrated the effectiveness of extracted semantic patterns.
