A Principal-Component-Analysis-Based Approach to Recognizing the Postures of Human Silhouettes

• Main Authors: Jiun-Liang Chen, 陳俊良
• Other Authors: Chin-Shyurng Fahn
• Format: Others
• Language: zh-TW
• Published: 2003
• Online Access: http://ndltd.ncl.edu.tw/handle/55524967456016448274

碩士 === 國立臺灣科技大學 === 自動化及控制研究所 === 91 === In this thesis, we present a principal component analysis (PCA) method to recognize body postures in real-time, which produces an eigenspace as our recognition model since a body posture in a two-dimensional image generally has a fixed shape and silhouette. Many different postures mainly made by various hands’ swing are used for identification. To get more complete foreground, the finding of a threshold for separating foreground from background followed by a noise removal process are first performed in our recognition system. At the beginning, this system adopts a background subtraction method to obtain foreground which is further preprocessed. Background subtraction is the method that the current image is cut off to get a rough foreground by means of finding an appropriate threshold to binarize the image so as to figure out foreground and background. Subsequently, both opening and closing operations in mathematical morphology are employed to erase noises. After this, connected components are searched for constitating a complete foreground. Next, we determine the contour pixel of foreground, which are sorted in an order and then recorded.. To recognize body postures, we take these ordered data as our main training samples. In experiments, we demonstrate 15 distinct postures and every posture has 10 samples, that is to say, we have 150 training samples. At last, we use the PCA method to find the most important projection vectors, and the corresponding posture can be obtained from solving the smallest vector distance between the input and training images. So far, the experimental results reveal that our approach is efficient and effective to recognize the body postures.