Walker Recognition Without Gait Cycle Estimation

Most of gait recognition algorithms involve walking cycle estimation to accomplish signature matching. However, we may be plagued by two cycle-related issues when developing real-time gait-based walker recognition systems. One is accurate cycle evaluation, which is computation intensive, and the other is the inconvenient acquisition of long continuous sequences of gait patterns, which are essential to the estimation of gait cycles. These drive us to address the problem of distant walker recognition from another view toward gait, in the hope of detouring the step of gait cycle estimation. This paper proposes a new gait representation, called normalized dual-diagonal projections (NDDP), to characterize walker signatures and employs a normal distribution to approximately describe the variation of each subject's gait signatures in the statistical sense. We achieve the recognition of unknown gait features in a simplified Bayes framework after reducing the dimension of raw gait signatures based on linear subspace projections. Extensive experiments demonstrate that our method is effective and promising.