Human metrology for person classification and recognition

Human metrological features generally refers to geometric measurements extracted from humans, such as height, chest circumference or foot length. Human metrology provides an important soft biometric that can be used in challenging situations, such as person classification and recognition at a distance, where hard biometric traits such as fingerprints and iris information cannot easily be acquired. In this work, we first study the question of predictability and correlation in human metrology. We show that partial or available measurements can be used to predict other missing measurements. We then investigate the use of human metrology for the prediction of other soft biometrics, viz. gender and weight. The experimental results based on our proposed copula-based model suggest that human body metrology contains enough information for reliable prediction of gender and weight. Also, the proposed copula-based technique is observed to reduce the impact of noise on prediction performance. We then study the question of whether face metrology can be exploited for reliable gender prediction. A new method based solely on metrological information from facial landmarks is developed. The performance of the proposed metrology-based method is compared with that of a state-of-the-art appearance-based method for gender classification. Results on several face databases show that the metrology-based approach resulted in comparable accuracy to that of the appearance-based method. Furthermore, we study the question of person recognition (classification and identification) via whole body metrology. Using CAESAR 1D database as baseline, we simulate intra-class variation with various noise models. The experimental results indicate that given enough number of features, our metrology-based recognition system can have promising performance that is comparable to several recent state-of-the-art recognition systems. We propose a non-parametric feature selection methodology, called adapted k-nearest neighbor estimator, which does not rely on intra-class distribution of the query set. This leads to improved results over other nearest neighbor estimators (as feature selection criteria) for moderate number of features. Finally we quantify the discrimination capability of human metrology, from both individuality and capacity perspectives. Generally, a biometric-based recognition technique relies on an assumption that the given biometric is unique to an individual. However, the validity of this assumption is not yet generally confirmed for most soft biometrics, such as human metrology. In this work, we first develop two schemes that can be used to quantify the individuality of a given soft-biometric system. Then, a Poisson channel model is proposed to analyze the recognition capacity of human metrology. Our study suggests that the performance of such a system depends more on the accuracy of the ground truth or training set.

Name thesis:degree_name

PhD

Level thesis:degree_level

Dissertation

Discipline thesis:degree_discipline

Lane Department of Computer Science and Electrical Engineering

Year dc:date.available

2013

Author dc:creator

• Cao, Deng

Contributors dc:contributor

• Donald Adjeroh
• Elaine Eschen
• James Harner
• Arun Ross
• Cun-Quan Zhang

• Computer science
• Statistics
• Biostatistics

Identifier

https://researchrepository.wvu.edu/etd/336

OAI identifier oai:identifier

oai:researchrepository.wvu.edu:etd-1339