| （英） |
Gait recognition is a biometric technology that identifies individuals based on their walking posture and motion, and it holds significant potential for applications such as criminal investigations and security gate systems. However, in the wild scenarios, subjects often walk directly toward the camera and are frequently occluded by objects. These conditions can lead to improper normalization when existing gait recognition methods are applied, resulting in recognition failures. To address this issue, we propose a normalization refinement network based on Spatial Transformer Networks (STNs), aiming to mitigate the negative impact of normalization. Specifically, we assume ten occlusion patterns with varying occlusion ratios and directions, and conduct evaluations on both normalization refinement accuracy and recognition performance. This paper presents experimental results obtained using the OUMVLP dataset. Furthermore, we demonstrate the applicability of the proposed network to GREW, a gait database captured in the wild, and report its performance under unconstrained conditions. |