Training and feature-reduction techniques for human identification using anthropometry

We investigate the utility of ID anthropometric measurements as a biometric for human identification when the subject pose differs in probe and gallery data. Whereas previous studies simulated probe data by adding noise to 3D gallery data, prior to extracting 1D measurements, we use a large 3D full-body data set having multiple poses per subject. Our analysis of 27 measurements from 2,144 subjects reveals differences due to pose, sensor, and other sources-all of which degrade recognition accuracy if uncompensated. We develop new training methods that use small sets of training data to measure and compensate for these differences. The new methods enable rank-1 identification >95% using 27 features and as few as 20 training subjects. To characterize the relative utility of the features and to simplify the biometric system, we develop techniques to identify feature subsets that together achieve good recognition performance. The reduction techniques demonstrate rank-1 identification of 83% and 94% using just ten and fifteen features. Together, these results will guide the development of more effective, accurate, and efficient anthropometry-based recognition systems.