Date of Graduation
Statler College of Engineering and Mineral Resources
Lane Department of Computer Science and Electrical Engineering
Over the years, biometric systems have evolved into a reliable mechanism for establishing identity of individuals in the context of applications such as access control, personnel screening and criminal identification. However, recent terror attacks, security threats and intrusion attempts have necessitated a transition to modern biometric systems that can identify humans under unconstrained environments, in real-time. Specifically, the following are three critical transitions that are needed and which form the focus of this thesis: (1) In contrast to operation in an offline mode using previously acquired photographs and videos obtained under controlled environments, it is required that identification be performed in a real-time dynamic mode using images that are continuously streaming in, each from a potentially different view (front, profile, partial profile) and with different quality (pose and resolution). (2) While different multi-modal fusion techniques have been developed to improve system accuracy, these techniques have mainly focused on combining the face biometrics with modalities such as iris and fingerprints that are more reliable but require user cooperation for acquisition. In contrast, the challenge in a real-time networked biometric system is that of combining opportunistically captured multi-view facial images along with soft biometric traits such as height, gait, attire and color that do not require user cooperation. (3) Typical operation is expected to be in an open-set mode where the number of subjects that enrolled in the system is much smaller than the number of probe subjects; yet the system is required to generate high accuracy.;To address these challenges and to make a successful transition to real-time human identification systems, this thesis makes the following contributions: (1) A score-based multi- modal, multi-sample fusion technique is designed to combine face images acquired by a multi-camera network and the effectiveness of opportunistically acquired multi-view face images using a camera network in improving the identification performance is characterized; (2) The multi-view face acquisition system is complemented by a network of Microsoft Kinects for extracting human anthropometric features (specifically height, shoulder width and arm length). The score-fusion technique is augmented to utilize human anthropometric data and the effectiveness of this data is characterized. (3) The performance of the system is demonstrated using a database of 51 subjects collected using the networked biometric data acquisition system.;Our results show improved recognition accuracy when face information from multiple views is utilized for recognition and also indicate that a given level of accuracy can be attained with fewer probe images (lesser time) when compared with a uni-modal biometric system.
Sankar, Sriram, "Effectiveness of Multi-View Face Images and Anthropometric Data In Real-Time Networked Biometrics" (2013). Graduate Theses, Dissertations, and Problem Reports. 5000.