Abstract
In this thesis, we introduce a real time people tracking detection and posture analysis under frequency domain and use a principal component analysis (PCA) model to generate an eigenspace. In general, the shape and silhouette of 2-D projection of human postures is changeless and it is a cue for us to determine what the person is doing. The postures we want to classify are stand, sit, bend, raise-hand, lying and walking. For these postures, we want to recognize each of them in spite of tiny variance of contour. In the whole system, in order to obtain a more clear and complete foreground silhouette, a method of background of maintenance and preprocessing of filtering noise is included.In the beginning of our system, there is a method of background subtraction and preprocessing, which include labeling of connected components, shape remover, morphology opening and closing. Then trace the contour pixels with a start point and sort them in order. With the ordered contour pixels, we can apply Fourier Descriptor (F.D.) on them to analysis the frequency component of our defined posture. And then estimate the similarity between F.D. of current image and the pre-stored data by our PCA models. With finding the maximum similarity, we can classify what the posture now is belonging to. The PCA model is used to generate an eigenspace by training each of the same posture of different people. And then we can recognize each posture according to these trained data.Under the condition and ability the hardware could provide, our system will analyze each frame of video input within about 1/15 second and achieve a near-real-time system.