HUMAN APPEARANCE MODELING IN VISUAL SURVEILLANCE

We present an appearance model for establishing correspondence between tracks of people which may be taken at different places, at different times or across different cameras. Illumination insensitive color features, i.e., RGB rank feature and brightness-color feature are used. Path-length feature is added for structural information and invariance to motion and pose. The appearance model is constructed by kernel density estimation. Kullback-Leibler distance measures the similarity between the models. To further exploit the information in video sequence, key frame selection method and online hierarchical clustering algorithm are proposed to construct appearance model from video. Key frame selection use the frames with large information gain to represent the appearance model. Online hierarchical clustering algorithm condense the model into a few clusters in the framework of our appearance model. Experimental results demonstrate the important role of the path-length feature in the appearance model and the effectiveness of the proposed appearance model and matching method