Scene Detection from High Definition MPEG Compressed Video Using Two Dimensional Anti-Dictionary

2014 Fiscal Year Final Research Report

• Project/Area Number: 24500110
• Research Category: Grant-in-Aid for Scientific Research (C)
• Allocation Type: Multi-year Fund
• Section: 一般
• Research Field: Media informatics/Database
• Research Institution: The University of Electro-Communications
• Principal Investigator: MORITA Hiroyoshi 電気通信大学, その他の研究科, 教授 (80166420)
• Co-Investigator(Kenkyū-buntansha): OTA Takahiro 長野県工科短期大学校, 情報技術科, 准教授 (60579001)
• Project Period (FY): 2012-04-01 – 2015-03-31
• Keywords: 反辞書 / オートマトン / 確率モデル / MPEG2 / シーン検出 / ビデオ解析

Outline of Final Research Achievements

We propose a method for analyzing video scenes directly from MPEG compressed video data without decoding it to pixel domains, and apply it to football programs to detect highlight-event scenes like corner kicks or goal shots. The proposed method achieves 86 % recall rate and 90 % precision rate in our experiments. In the proposed method, the information of macro block type (MBT) which indicates which type of coding methods is used for encoding each macroblock in a video frame, plays an important role. As a useful tool to analyze MBT pattern which is an array of MBT's in a frame, we propose a two dimensional antidictionary coding method. A two dimensional antidictionary of a given pattern P is a set of minimum forbidden pattern which does not appear in P but whose sub-patterns appear in P. In this research, we establish a method for construction a two dimensional antidictionary and consider applying it to video analysis.

Free Research Field

情報理論，画像工学