Deterministic Learning for Human Gait Recognition

In this paper, we investigate the problem of human gait recognition based on temporal data sequences by utilizing deterministic learning theory. We employ joint-angle trajectories of lower limbs for gait recognition. The joint angle data generated from human gait locomotion is chosen as the temporal data since it represents the dynamic characteristics of human gait and contains lots of valuable information for gait recognition. What's more, the joint angle trajectories are periodic or periodic-like (recurrent) which makes the radial basis function (RBF) network easily satisfy the partial persistence of excitation (PE) condition. Firstly, discrete-time joint angle data obtained by motion-capture equipment or image-processing algorithms forms the temporal data sequences generated from human gait locomotion, locally-accurate approximation of the underlying gait system dynamics is achieved by using RBF networks. We then prove the convergence of the approximation error and related parameters. Consequently, the joint angle data sequences can effectively represent human gait locomotion by using the knowledge of approximated gait dynamics which is kept in constant RBF networks. Finally, similarity definition for temporal gait data sequences generated from different persons or from different status of one person is given, and a method for recognition of gait temporal data sequences is proposed. We use less complicated simulation examples of compass-like biped robots gait recognition to demonstrate the effectiveness of our schemes.