Autonomous surveillance of a dynamic, multi-subject environment: an implementation for automated facial recognition

The paper presents a novel reconfigurable surveillance system that incorporates multiple active-vision cameras aimed for improved automated facial recognition of a subject as he/she travels through a cluttered environment. The proposed system utilizes cameras with both rotational and translational capabilities in order to avoid occlusions and gain an advantaged viewpoint. Further performance improvements are achieved through selection and fusion of data from multiple redundant cameras. A novel facial-recognition algorithm is developed for recognition of subjects from varying viewing angles by utilizing and fusing databases of frontal and profile images. It is shown through experiments conducted on the FERET database that the proposed facial recognition algorithm reduces sensitivity to viewing angle over conventional methods. Experiments conducted with an experimental surveillance system show that recognition performance can be tangibly improved through the use of multiple active-vision cameras coupled with an effective dispatching strategy.