Gait-Based Human Recognition by Classification of Cyclostationary Processes on Nonlinear Shape Manifolds

We study the problem of analyzing and classifying human gait by modeling it as a stochastic process on a shape space. We consider gait as a evolution of human silhouettes as seen in video sequences, and focus on their shapes. More specifically, we define a shape space of planar, closed curves and model a human gait as a stochastic process on this space. Due to the periodic nature of human walk, this process is naturally constrained to be cyclostationary, that is, its mean path is assumed to be cyclic. We compare two subjects using a metric that quantifies differences between average gait cycles of each subject. This computation uses several tools from differential geometry of the shape space, including computation of geodesics, estimation of means of observed shapes, interpolation between observed shapes, and temporal registration of two gait cycles. Finally, we apply a nearest-neighbor classifier, using the gait metric, to perform human recognition, and present results from an experiment involving 26 subjects.