Facial Re-identification on Non-overlapping Cameras and in Uncontrolled Environments

Abstract

Face re-identification is an essential task in automatic video surveillance where the identity of the person is known previously. It aims to verify if other cameras have observed a specific face detected by a camera. However, this is a challenging task because of the reduced resolution, and changes in lighting and background available in surveillance video sequences. Furthermore, the face to get re-identified suffers changes in appearance due to expression, pose, and scale. Algorithms need robust descriptors to perform re-identification under these challenging conditions. Among various types of approaches available, correlation filters have properties that can be exploited to achieve a successful re-identification. Our proposal makes use of this approach to exploit both the shape and content of more representative facial images captured by a camera in a field of view. The resulting correlation filters can characterize the face of a person in a field of view; they are good at discriminating faces of different people, tolerant to variable illumination and slight variations in the rotation (in/out of plane) and scale. Further, they allow identifying a person from the first time that has appeared in the camera network. Matching the correlation filters generated in the field of views allows establishing a correspondence between the faces of the same person viewed by different cameras. These results show that facial re-identification under real-world surveillance conditions and biometric context can be successfully performed using correlation filters adequately designed.