Human identification based on Gait Manifold

Gait-based pedestrian identification has important applications in intelligent surveillance. From anatomical viewpoint, the physical uniqueness of human gait is physiological discriminative of individuals. Therefore, in theory, like fingerprint and face, gait is used as a biometric for pedestrian identification. However, gait-based pedestrian identification still faces multiple challenges due to a vast diversity of walking conditions and complex acquisition environment during data collection. In this paper, through combining the nonlinear dimensionality reduction by using gait manifold and the temporal feature of gated recurrent unit (GRU) together, we propose a novel gait-based pedestrian identification framework. Firstly, we design a temporal enhancement module to construct a series of frame-by-frame gait trend energy images (ff-GTEIs), which represents spatiotemporal gait characteristics and does not reduce the number of samples. Secondly, a supervised locally linear embedding (LLE) dimensionality reduction scheme is proposed, which generates a low-dimensional gait manifold for each pedestrian and transforms all ff-GTEIs into corresponding gait manifold space. Thirdly, a new pedestrian identification network based on residual GRU is proposed, which is able to identify a person by comprehensively considering the similarity between its gait and corresponding gait manifold. Finally, a series of comparative experiments are carried out based on well-known gait datasets, the experimental results show that the proposed framework for pedestrian recognition in this paper exceeds most existing methods, and has achieved an average correct recognition rate 97.4% and 99.6% based on the open-accessed gait dataset CASIA Dataset B and OU-ISIR LP dataset, respectively.