Date of Graduation
Statler College of Engineering and Mineral Resources
Lane Department of Computer Science and Electrical Engineering
Arun A. Ross
In biometrics-based surveillance and de-duplication applications, the system commonly determines if a given individual has been encountered before. In this dissertation, these applications are viewed as specific instances of a broader class of problems known as Anonymous Identification. Here, the system does not necessarily determine the identity of a person; rather, it merely establishes if the given input biometric data was encountered previously. This dissertation demonstrates that traditional biometric evaluation measures cannot adequately estimate the error rate of an anonymous identification system in general and a de-duplication system in particular. In this regard, the first contribution is the design of an error prediction model for an anonymous identification system. The model shows that the order in which individuals are encountered impacts the error rate of the system. The second contribution - in the context of an identification system in general - is an explanatory model that explains the relationship between the Receiver Operating Characteristic (ROC) curve and the Cumulative Match Characteristic (CMC) curve of a closed-set biometric system. The phenomenon of biometrics menagerie is used to explain the possibility of deducing multiple CMC curves from the same ROC curve. Consequently, it is shown that a "good'' verification system can be a "poor'' identification system and vice-versa.;Besides the aforementioned contributions, the dissertation also explores the use of gait as a biometric modality in surveillance systems operating in the thermal or shortwave infrared (SWIR) spectrum. In this regard, a new gait representation scheme known as Gait Curves is developed and evaluated on thermal and SWIR data. Finally, a clustering scheme is used to demonstrate that gait patterns can be clustered into multiple categories; further, specific physical traits related to gender and body area are observed to impact cluster generation.;In sum, the dissertation provides some new insights into modeling anonymous identification systems and gait patterns for biometrics-based surveillance systems.
DeCann, Brian Matthew, "Modeling Errors in Biometric Surveillance and De-duplication Systems" (2014). Graduate Theses, Dissertations, and Problem Reports. 466.