| Summary: | 碩士 === 國立中興大學 === 電機工程學系 === 93 === This thesis proposes two human posture analysis methods, one is human body posture recognition by neural fuzzy network, and the other is human posture estimation by silhouette. For posture recognition, four kinds of main body postures, including standing, sitting, lying, and bending, are recognized. We use a moving object segmentation algorithm to distinguish the human body and background from a sequence images. After the human body is successfully segmented, we use a sequence of image processing algorithms, including median filter and morphological operation, to obtain a complete silhouette. We project the silhouette onto horizontal and vertical axes, respectively, and find Discrete Fourier Transform of each. Significant Fourier transform values together with length-width ratio of the silhouette are used as features. Recognizer is designed by a neural fuzzy network. Experimental results show that we can recognize the four postures with a high accuracy. Simulations on applying the recognition approach to the detection of the emergency condition that one suddenly falls down and can not stand up are also performed.
In posture estimation, our objective is to locate significant body points, including head, tips of hand, and tips of feet. Based on the silhouette, we compute center of gravity (COG) of human body and boundary contour. By computing the distance between COG and pixels in the contour, we obtain a curve of distance. Concave points in the distance curve are located and regarded as candidates of the significant body points. Based on orientation of the body, body structure, and curvature of candidates, we first locate the hand followed by the location of tips of feet and tips of hand. Experiments show that the proposed approach can recognize significant points of most postures.
|
|---|
