Appearance-based person recognition using color/path-length profile

Recognizing a person is a basic process of video understanding. In this paper, we present a method for recognizing (matching) persons based on their overall extrinsic appearance, regardless of their (upright) pose. Appearance is that of their visible clothing and bodies seen in silhouette obtained by background subtraction. The method of appearance recognition uses kernel estimation of probabilities associated with color/path-length profiles and uses Kullback–Leibler distance to compare such profiles with possible models. We show that with suitable normalization of color variables our method appears to be robust under conditions varying viewpoints, complex illumination, and multiple cameras. Our method is also useful to detect changes in appearance, for instance caused by carried packages. When there are more than one profiles to match in one frame, we adopt multiple matching algorithm enforcing 1 to 1 constraint to improve the performance.