| In this thesis, the methods of video analysis for visual surveillance are studied in order to develop an algorithm that detects, tracks and recognizes people at an entrance to a room. To detect people, background subtraction and silhouette analysis are used. To track and recognize people, the colour of a person's clothing is used.; Each background pixel is modeled using a single Gaussian distribution. This model is updated using the corresponding pixel intensity and the variation in intensity of the pixel from one image in the sequence to the next. A pixel in the current image is considered foreground if its intensity is not within a given number of standard deviations of the corresponding background pixel model.; Silhouette analysis is used to determine if a foreground region represents a person. The curvature of the top portion of a silhouette is analyzed to determine if it conforms to the shape of a person's head. Then the percentage of foreground pixels in a region of size proportional to the width of a detected head and immediately below the detected head is determined. If this value is large enough then a person is detected.; To track and recognize people, colour histograms extracted from each person's clothing in a luminance and perceptually uniform chrominance space are compared using the Earth Mover's Distance. Clothing colour is extracted from the same region defined for person detection. The goal is to determine whether a person has entered or exited the room and to associate a sequence showing a person leaving the room with the previously recorded sequence showing that same person entering. |
