Cross-Attention Framework for Gait Recognition Under Clothing Variations

Gait recognition is a non-contact biometric technology used for behavior tracking and condition monitoring in various settings like facilities, nursing homes, and hospitals. However, traditional silhouette-based methods are susceptible to factors like clothing, belongings, and lighting changes, leading to reduced performance in real-world environments. To address this, we applied a deep learning-based clothing augmentation technique to simulate diverse clothing conditions, creating a test set that mirrors real street environments. The proposed model enhances gait recognition by separating the upper and lower body using skeleton-based structural information, tracking joint displacement between frames, and applying differential weights to areas of interest. Additionally, Flash Attention, an advanced fusion structure, was used to model feature interactions accurately and minimize information loss. Experiments on GREW, CCVID, and a custom-built real-world test set showed that the proposed model maintained robust recognition performance under varying clothing and background conditions. These results demonstrate that the model can be effectively applied in real-world scenarios.