In this paper we address the problem of human gait recognition from a robust identification and model (in)validation prospective. The main idea is to apply dimensionality reduction technique to extract the spatio-temporal information by mapping the gait silhouette sequence to a low dimensional time sequence, which is considered as the output of a linear time invariant (LTI) system. A class of gaits is associated to a nominal discrete LTI system which has a periodic impulse response and is identified by robust identification approach. Correspondingly, gait recognition can be formulated as measuring the difference between the models representing different gait sequences. Our approach provides an efficient way to extract, to model shape-motion information of gait sequence, and to measure the difference between gait sequence models which is robust to gait cycle localization, gross appearance variation, and time scaling. These results are illustrated with practical examples on popular gait databases.