Automatic gait recognition via model-based moving feature analysis

Cunado, David (1999) Automatic gait recognition via model-based moving feature analysis. University of Southampton, Doctoral Thesis.

Abstract

Modern day society requires ever improving means of person authentication and recognition. Biometrics, measurements taken from a human body for the sole purpose of identification/recognition, have increased in variety, reliability and performance due to advances in technology. Even with this large research area, studies on using gait as a biometric have appeared only recently. Gait is an attractive biometric as its measurement can be remote, without subject contact. Also, in many applications of person identification particularly those involving crime, many established biometrics can be obscured. Since people need to walk, their gait is usually apparent. This thesis examines the potential of gait as a form of person identification using computer vision techniques. A feature-based approach is developed, extracting a biometric measure with a clear analytic justification using a model based on medical studies. A gait signature was derived from this biometric measure, the hip rotation pattern.

A preliminary study was performed in which extant computer vision techniques were used to used to track the upper legs of a subject through a sequence of images. Using the Hough transform for lines, the inclination of the lines that best represented the thigh in each image of a sequence was collated. The collected data was fitted to a polynomial using least-squares analysis, producing the hip rotation pattern for the subject. The magnitude and phase component of the hip motion was found using the Detector Fourier transform, and a gait signature was formed from these components. Experimentation was performed on a database of ten subjects, and classification using k-nearest classifier generated an improved correct classification rate when using the phase-weighted Fourier magnitude information (80% for k= 1 and 90% for k = 3) than when using magnitude information alone (40% for k = 1 and 50% k = 3).

The techniques developed in the preliminary study produced encouraging results, demonstrating that it did indeed appear possible to recognise people from their gait using computer vision techniques. Although a new approach had been developed, there was limited efficiency due to several strategic problems. Frame-by-frame analysis of the image sequence produced a temporal feature extraction method that is very susceptible to noise and therefore prone to experience missing features. The collation of hip inclinations when forming the hip rotation patterns was open to inconsistency. Also, the use of a polynomial to model the motion of the thigh was inconsistent with the notion of gait as a periodic signal. As such, a new feature extraction technique was required to eliminate these shortfalls, aiming for automatic signature generation.

This record has no associated files available for download.

Contributors

Download statistics