Model-based 3D gait biometrics

Abstract

There have as yet been few gait biometrics approaches which use temporal 3D data. Clearly, 3D gait data conveys more information than 2D data and it is also the natural representation of human gait perceived by human. In this paper we explore the potential of using model-based methods in a 3D volumetric (voxel) gait dataset. We use a structural model including articulated cylinders with 3D Degrees of Freedom (DoF) at each joint to model the human lower legs. We develop a simple yet effective model-fitting algorithm using this gait model, correlation filter and a dynamic programming approach. Human gait kinematics trajectories are then extracted by fitting the gait model into the gait data. At each frame we generate a correlation energy map between the gait model and the data. Dynamic programming is used to extract the gait kinematics trajectories by selecting the most likely path in the whole sequence. We are successfully able to extract both gait structural and dynamics features. Some of the features extracted here are inherently unique to 3D data. Analysis on a database of 46 subjects each with 4 sample sequences, shows an encouraging correct classification rate and suggests that 3D features can contribute even more.

Gait recognition based on joint distribution of motion angles

“…In [13,14], based on the appearance and view point, Kale et al presented the binarization contour as the feature using the Dynamic Time Warping (DTW) to deal with the speed changes in the process of walking, and strengthen the fault tolerance of original data. In addition, Kale et al [9] use the width of the outer contour of the binarized silhouette as the image feature and built a HMM model to distinguish the dynamic features of the gait. Hu et al [15] also built a HMM model for gait recognition where they took the local binary pattern descriptor as the motion feature.…”

“…In the model-based approaches, the prior information or a known model is needed to fit human gait. Though the model-based method is more complex, it has some advantages such as immunity to noise [9]. The model-based approaches assume a priori model to match the data extracted from video, and features correspondence is automatically achieved.…”

A New View-Invariant Feature for Cross-View Gait Recognition

“…The (2) can be reformulated as (3) Thus, given a corrupted GTI observed from a certain view, we need to recover the low-rank texture , the sparse error matrix , and the domain transformation . This naturally leads to the optimization problem as (4) where denotes the number of nonzero entries in . That is, we aim to find with the lowest possible rank and with the fewest possible nonzero entries that agree with the observed GTI up to the domain transformation .…”

“…Methods in the first category [2]- [4] are to construct 3-D gait information through multiple calibrated cameras. Then, 2-D gait information from any required view is reconstructed from the 3-D gait information.…”

Marionette mass-spring model for 3D gait biometrics

“…This model usually is used to tract the object data in order to recover the body pose and obtain the motion trajectories. Another 3D recognition study [1] used a 3D model based on using the match between volumetric cylinders and the voxel data, by correlation filter and dynamic programming.…”