近年來行為分析在多媒體影像處理的問題中,扮演著很重要的角色,同時有許多實際應用於生活之中,例如:社區監控、視訊監控、遠距醫療監控、異常行為偵測…等。一般的行為分析研究限制於固定視角,但於現實環境當中,人類行為不會只在預設的視角下被察覺。因此,本篇論文著重於多視角環境下使用單一攝影機之人類行為的分析與辨識。 本系統將多視角之行為分析轉換至單一視角下處理。先使用姿勢樣板比對找出最相似的姿勢樣板,再將樣板比對結果轉換至指定視角當中同姿勢的樣板。利用樣板之間的相關性與樣版轉換機率的輔助,使得指定視角中的樣板轉換有準確的結果。將此連續的樣板存入統計陣列之中,將此陣列與資料庫之行為陣列進行辨識,可分析出此視訊之中發生的行為。 由實驗結果得知,我們的系統可以適用於多視角環境下之行為分析系統,並且有良好的表現。
This paper presents a new behavior classification system that can be used to analyze human movements from any views and directly form videos. Technically, if a person is observed from other views, his appearances will change significantly. To freely recognize his behaviors, traditional methods tend to adopt 3-D data for behavior modeling and analysis. However, its inherent correspondence process is very time-consuming and will make it inappropriate for real time applications. To tackle this problem, this thesis proposes a novel human representation scheme for recognizing human behaviors from any views. In this scheme, a novel view alignment method is first proposed for mapping each action sequence (captured from any views) to a fixed view. To achieve this mapping, the spatial and temporal features should be first extracted from each action sequence. To extract the spatial feature, the central contexts of each posture are then extracted through a triangulation technique. Then, a set of key postures is selected and built for converting an action to a symbol string. To reduce the converting errors, a transitional probably table is built for recording the possibility of one posture transferring to another posture. With the table and the central context feature, each action sequence can be then aligned to a fixed view and then represented by an action matrix. After that, matching two action sequences from arbitrary views will become a single-view matrix matching problem. Then, the Viterbi algorithm is used for aligning two action sequences and then classifying them to different behavior types. Experimental results prove that the proposed method is a robust and accurate tool for human movement analysis from any views.