SpecGait: spatial coordinate aggregation and dynamic-static decoupling network for in-the-wild gait recognition

Abstract

As a non-contact long-distance biometric identification technology, gait recognition has significant advantages in the field of public safety. However, when landing in the field scene, it is disturbed by covariates, and there is a problem of gait feature extraction distortion. The existing methods are difficult to capture the spatial position information of the human body, eliminate the interference such as replacement, and have weak anti-contour deformation ability. To this end, this paper proposes a SpecGait spectrum-space decoupled gait recognition network, designs SFA-Block to enhance spatial perception from bidirectional coded position information, and innovates DSD-Module to split dynamic and static features and purify core gait information by referring to signal decomposition ideas. It is verified by Gait3D and BUAA-Duke-Gait datasets that the network recognition accuracy and anti-interference performance are better than the existing algorithms, which confirms the practical value of the proposed strategy.