Authors: Dustin A. Bruening; Charles D. Goodyear; David R. Bowden; Rebecca E. Barone
Addresses: Department of Exercise Sciences, Brigham Young University, Provo, Utah, USA ' Infoscitex: a DCS Company, Dayton, Ohio, USA ' Air Force Research Laboratories, Dayton, Ohio, USA ' Infoscitex: a DCS Company, Dayton, Ohio, USA
Abstract: Automated gender recognition from whole body images is a challenging problem with multi-disciplinary utility. A greater understanding of potential feature components (e.g., anthropometry, movement, etc.) may help future feature selection algorithms better target effective features, reduce feature complexity, and increase algorithm generalisability. In this study we evaluated the potential of static anthropometric measurements for gender recognition. Utilising a large 3D body scan repository, we first captured novel measurements directly relevant to computer vision applications, and used these to create biologically guided feature sets. Linear discriminant analysis was used to classify gender across specific demographics to additionally evaluate the potentially confounding influences of race, age, and obesity. The effects of view angle were also preliminarily analysed. Classification results showed greater accuracy in the frontal plane than the sagittal plane, with models reaching 99% and 96% accuracy, respectively. Feature rankings and correlations are presented and discussed in relevance to future algorithms.
Keywords: soft biometrics; sex recognition; anthropomorphics; linear discriminant analysis; LDA; feature selection; feature correlation; gender recognition; anthropometric features; whole body images; computer vision; demographics; race; age; obesity; view angle.
International Journal of Biometrics, 2015 Vol.7 No.4, pp.354 - 372
Received: 13 Apr 2015
Accepted: 01 Dec 2015
Published online: 26 Apr 2016 *