Semester

Summer

Date of Graduation

2012

Document Type

Thesis

Degree Type

MS

College

Statler College of Engineering and Mineral Resources

Department

Lane Department of Computer Science and Electrical Engineering

Committee Chair

Natalia A. Schmid.

Abstract

Biometric systems have been typically designed to operate under controlled environments based on previously acquired photographs and videos. But recent terror attacks, security threats and intrusion attempts have necessitated a transition to modern biometric systems that can identify humans in real-time under unconstrained environments. Distributed camera networks are appropriate for unconstrained scenarios because they can provide multiple views of a scene, thus offering tolerance against variable pose of a human subject and possible occlusions. In dynamic environments, the face images are continually arriving at the base station with different quality, pose and resolution. Designing a fusion strategy poses significant challenges. Such a scenario demands that only the relevant information is processed and the verdict (match / no match) regarding a particular subject is quickly (yet accurately) released so that more number of subjects in the scene can be evaluated.;To address these, we designed a wireless data acquisition system that is capable of acquiring multi-view faces accurately and at a rapid rate. The idea of epipolar geometry is exploited to get high multi-view face detection rates. Face images are labeled to their corresponding poses and are transmitted to the base station. To evaluate the impact of face images acquired using our real-time face image acquisition system on the overall recognition accuracy, we interface it with a face matching subsystem and thus create a prototype real-time multi-view face recognition system. For front face matching, we use the commercial PittPatt software. For non-frontal matching, we use a Local binary Pattern based classifier. Matching scores obtained from both frontal and non-frontal face images are fused for final classification. Our results show significant improvement in recognition accuracy, especially when the front face images are of low resolution.

Share

COinS