Modeling and Recognizing Human Activities from Video

• Item type: SIG Technical Reports(1)
• 公開日: 2009-06-02
• タイトル: Modeling and Recognizing Human Activities from Video
• タイトル / 言語: en
• タイトル: Modeling and Recognizing Human Activities from Video
• 言語: eng
• キーワード / 主題Scheme: Other
• キーワード / 主題: D論セッション2
• 資源タイプ / 資源タイプ識別子: http://purl.org/coar/resource_type/c_18gh
• 資源タイプ: technical report
• 著者所属: University of Electro-Communications / University of Tokyo
• 著者所属: University of Tokyo
• 著者所属(英): en
• 著者所属(英): University of Electro-Communications / University of Tokyo
• 著者所属(英): en
• 著者所属(英): University of Tokyo
• 著者名: KrisM.Kitani; Yoichi, Sato
• 著者名(英): Kris, M.Kitani; Yoichi, Sato
• 論文抄録 / 内容記述タイプ: Other
• 論文抄録 / 内容記述: This paper presents a complete computational framework for discovering human actions and modeling human activities from video, to enable intelligent computer systems to effectively recognize human activities. A bottom-up computational framework for learning and modeling human activities is presented in three parts. First, a method for learning primitive actions units is presented. It is shown that by utilizing local motion features and visual context (the appearance of the actor, interactive objects and related background features), the proposed method can effectively discover action categories from a video database without supervision. Second, an algorithm for recovering the basic structure of human activities from a noisy video sequence of actions is presented. The basic structure of an activity is represented by a stochastic context-free grammar, which is obtained by finding the best set of relevant action units in a way that minimizes the description length of a video database of human activities. Experiments with synthetic data examine the validity of the algorithm, while experiments with real data reveals the robustness of the algorithm to action sequences corrupted with action noise. Third, a computational methodology for recognizing human activities from a video sequence of actions is presented. The method uses a Bayesian network, encoded by a stochastic context-free grammar, to parse an input video sequence and compute the posterior probability over all activities. It is shown how the use of deleted interpolation with the posterior probability of activities can be used to recognize overlapping activities. While the theoretical justification and experimental validation of each algorithm is given independently, this work taken as a whole lays the necessary groundwork for designing intelligent systems to automatically learn, model and recognize human activities from a video sequence of actions.
• 論文抄録(英) / 内容記述タイプ: Other
• 論文抄録(英) / 内容記述: This paper presents a complete computational framework for discovering human actions and modeling human activities from video, to enable intelligent computer systems to effectively recognize human activities. A bottom-up computational framework for learning and modeling human activities is presented in three parts. First, a method for learning primitive actions units is presented. It is shown that by utilizing local motion features and visual context (the appearance of the actor, interactive objects and related background features), the proposed method can effectively discover action categories from a video database without supervision. Second, an algorithm for recovering the basic structure of human activities from a noisy video sequence of actions is presented. The basic structure of an activity is represented by a stochastic context-free grammar, which is obtained by finding the best set of relevant action units in a way that minimizes the description length of a video database of human activities. Experiments with synthetic data examine the validity of the algorithm, while experiments with real data reveals the robustness of the algorithm to action sequences corrupted with action noise. Third, a computational methodology for recognizing human activities from a video sequence of actions is presented. The method uses a Bayesian network, encoded by a stochastic context-free grammar, to parse an input video sequence and compute the posterior probability over all activities. It is shown how the use of deleted interpolation with the posterior probability of activities can be used to recognize overlapping activities. While the theoretical justification and experimental validation of each algorithm is given independently, this work taken as a whole lays the necessary groundwork for designing intelligent systems to automatically learn, model and recognize human activities from a video sequence of actions.
• 書誌レコードID / 収録物識別子タイプ: NCID
• 書誌レコードID / 収録物識別子: AA11131797
• 書誌情報: 研究報告コンピュータビジョンとイメージメディア（CVIM）; 巻 2009-CVIM-167, 号 3, p. 1-16, 発行日 2009-06-02
• Notice: SIG Technical Reports are nonrefereed and hence may later appear in any journals, conferences, symposia, etc.
• 出版者 / 言語: ja
• 出版者: 情報処理学会