Model-based 3D gait biometrics

Abstract

Gait biometrics has attracted increasing interest in the computer vision and machine learning communities because of its unique advantages for recognition at distance. However, there have as yet been few gait biometric approaches which use temporal three-dimensional (3D) data. Clearly, 3D gait data conveys more information than 2D gait data and it is also the natural representation of human gait as perceived by humans. The University of Southampton has created a multi-biometric tunnel using twelve cameras to capture multiple gait images and reconstruct them into 3D volumetric gait data. Some analyses have been done using this 3D dataset mainly to solve the view dependent problem using model-free silhouette-based approaches. This thesis explores the potential of model-based methods in an indoor 3D volumetric gait dataset and presents a novel human gait features extraction algorithm based on marionette and mass-spring principles.

We have developed two different model-based approaches to extract human gait kinematics from 3D volumetric gait data. The first approach used a structural model of a human. This model contained four articulated cylinders and four joints with two degrees of rotational freedom at each joint to model the human lower legs. Human gait kinematic trajectories were extracted by fitting the gait model to the gait data. We proposed a simple yet effective model-fitting algorithm using a correlation filter and dynamic programming.

To increase the fitting performance, we utilized a genetic algorithm on top of this structural model. The second approach was a novel 3D model-based approach using a marionette-based mass-spring model. To model the articulated human body, we used a stick-figure model which emulates marionette's motion and joint structure. The stick-figure model had eleven nodes representing the human joints of head, torso and lower legs. Each node was linked with at least one other node by spring. The voxel data in the next frame had a role as an attractor which able to generate forces for each node and then iteratively warp the model into the data. This process was repeated for successive frames.

Our methods can extract both structural and dynamic gait features. Some of the extracted features were inherently unique to 3D gait data such as footprint angle and pelvis rotation. Analysis on a database of 46 subjects shows an encouraging correct classification rate up to 95.1% and suggests that model-based 3D gait analysis can contribute even more in gait biometrics.

More information

Identifiers

ORCID for Mark S. Nixon: orcid.org/0000-0002-9174-5934

Catalogue record

Export record