標題: 以輪廓為基礎之人類行為分析
Silhouette-Based Human Behavior Analysis
作者: 黃雅楠
Alan Huang
廖弘源
資訊科學與工程研究所
關鍵字: 行為;輪廓;追蹤;曲度;混合高斯模型;Human;Behavior;Silhouette;Tracking;Human Body;Curvature;GMM
公開日期: 2004
摘要: 在這篇論文裡,我們發展了一套具有分析人類行動能力的即時視覺監視系統(real-time visual surveillance system)。設備為一支彩色單眼錄影機;環境參數為靜止不動的背景(stationary background)。在一開始,系統會結合傳統的電腦追蹤演算法-背景相減法及混合高斯模型(GMM),將出現在鏡頭裡的目標物偵測出來,並萃取出其輪廓。接著,系統使用形狀及幾何分析所獲得的特徵值,成功的將人體各部位區分開成尚未定義的區塊。並利用我們所發展的階層式形狀統計相似演算法(HSSS)標明被區分開的區域(頭、手、軀體…等)。當系統成功地將以上區塊分辯出來,便會估算出被觀察者身體各部位的物理特徵值,例如:型體重心、主軸角度、長寬比例…等。最後再利用這些估算值,與事先建立好的資料庫作比對。此資料庫是以動作(actions)為基礎所建立的。藉此系統可以自動地監視被觀察者以及適時地發出警告訊息。此系統可以20~25Hz的速度、240x160的解析度在Pentium-M 1600MHz的PC上作用。
In this thesis, we develop a real-time visual surveillance system for human behavior analysis. It operates on monocular color-scale video imagery with a stationary background scene. At the first step in the system process, it extracts the silhouette of the target object by traditional video tracking method, background subtraction combined with Gaussian Mixture Model ( GMM ). And furthermore, it detects the contour of the object silhouette. At the second step, the system employs a combination of shape analysis and geometry analysis on the contour to decompose the detected silhouette to several undefined parts(unlabeled body parts). After the decomposition process, it labels each the separated part (head、torso、hands、feet)by the use of our hierarchical statistical-shape-similarity algorithm ( HSSS ). As the above steps have been processed successfully, the last step in our system is to extract local features of the detected body part(orientation、centroid. . . etc), and the global features of the entire silhouette(aspect ratio、 block density. . . etc), and then these features can be used to guide the high-level human behavior analysis. In the on-line behavior analysis process, an unknown sequence will be matched with the templates collected in our database. The database is established offline by the use of real video captures, which is a group of labeled reference sequence representing typical behaviors. In short, our system can detect the human body parts and classify the posture of human at individual imagery, then identify the event of a query sequence which involves human beings. It runs at 20~25Hz for 240 x 160 resolution images on a single Pentium-M 1600Mhz PC.
URI: http://140.113.39.130/cdrfb3/record/nctu/#GT009217634
http://hdl.handle.net/11536/74390
Appears in Collections:Thesis


Files in This Item:

  1. 763401.pdf
  2. 763402.pdf