Gait recognition based on 3D human body reconstruction and multi-granular feature fusion

Gait recognition is a crucial video-based biometric approach that allows for the identification of pedestrians from the motion of their walk over a distance without direct contact. Despite significant advances in this field, most existing approaches for gait recognition rely on silhouette sequence extraction, which can result in redundant information when the behavior of pedestrians changes, such as with the addition of coats or bags. To alleviate this, we propose an end-to-end gait recognition method based on 3D human body reconstruction to effectively remove this redundant information and generate compact, discriminative gait representations. Furthermore, to make full use of the spatial characteristics of pedestrians, we propose a multi-granular feature fusion module to model gait representations at multiple granularities. Our method is evaluated on the Outdoor-Gait and CASIA-B datasets and shows improved performance and robustness.