We propose a fast retrieving method for searching a video segment for a number of video streams to detect illegal video copies on Internet video site and to search the original video source at TV stations, video production companies and so on. We extract a few and a small number of feature vectors from feature points that express the point of local maximum/minimum value in the time sequence of audio power in video data. The relative values of the feature points are called as time-space location information. These features can be easily obtained compared with the representative features such as MFCC, and the processing time for matching becomes short because the number and the dimension of each feature vector are both small. We evaluated the proposed method for the accuracy and the computation time. Assuming a copyright holder of a movie searches the illegal copies for plenty of video streams, and we use broadcasted TV movie programs recorded in an HDD recorder for input data, and a 30sec-3min long segment in DVD for reference data. As the result, we could confirm that the proposed method completed all processes including the feature extraction and the matching process within the computation time of the feature extraction in [3-4], and the proposed method achieved 93.2% of F-measure in the video segment detection.%手元にあるビデオを用いて，動画サイト上の違法コピー検出やテレビ局等でのビデオソース検索等を目的として，ビデオ音響部の少数かつ低次元の時空間配置情報を特徴量として用いた高速な照合方式を提案する，ビデオ音響部のパワー情報から極大・極小値となる箇所を特徴点として抽出し，その相対値を時空間配置情報として特徴量とする.これらの特徴量はMFCCやLPCケプストラム等の特徴量と比べ簡単に計算でき，特徴点も少数かつ低次元になるため高速な照合を実現できる.著作権の所有者がビデオ群から実際に検出を行う事を想定して，DVDから得た音響情報と同一内容の放送番組を一旦録画（音）した環境の異なる音響情報を用いて，区間検出精度と計算時間について評価を行った.評価実験の結果，先行研究［3-4］の特徴量作成時間内に，提案方式は特徴量作成から照合までの処理を終え，F値93.2％で区間検出できること確認した.
展开▼
