Time-series Active Search: Instantaneous Searching of Audio and Video

摘要

This paper proposes a search method that can quickly detect and locate a desired audio or video segment in a long audio or video stream. The method, called time-series active search, is based on histo- gram modeling of audio or video features. Histogram modeling enables us to reduce the number of candidate matches between reference and input signals by approximately 100 to 500 times compared to exhaustive search, while guaranteeing the same retrieval accuracy. Experiments show that the pro- posed method can correctly retrieve a 15-s segment from a 6-h TV broadcast-quality recording within 2.3 s (for audio) or 1.0 s(for video), once the feature vectors have been extracted.