GLGAIT: ENHANCING GAIT RECOGNITION WITH GLOBAL-LOCAL TEMPORAL RECEPTIVE FIELDS FOR IN-THE-WILD SCENARIOS

Gait recognition has emerged as a promising non- intrusive human identification technology, capable of functioning from a distance without requiring active cooperation. While existing methods have shown considerable success in controlled laboratory environments, their performance often degrades in real-world scenarios. To address this challenge, we propose GLGait, a novel framework designed to improve gait recognition in the wild by establishing comprehensive temporal receptive fields. The core of GLGait is the Global-Local Temporal Module (GLTM), which integrates a Pseudo Global Temporal Self- Attention (PGTA) mechanism with temporal convolution operations. PGTA effectively captures pseudo-global temporal patterns with reduced computational overhead compared to traditional multi-head self-attention (MHSA), while temporal convolution enhances local temporal representations and aggregates them into a unified holistic temporal receptive field. Additionally, we introduce a Center-Augmented Triplet Loss (CTL), which minimizes intra-class distances and increases positive sample rep- resentation during training. Extensive evaluations on challenging in-the-wild datasets, such as Gait3D and GREW, demonstrate that GLGait achieves state-ofthe-art performance, offering a robust and efficient solution for gait recognition in unconstrained settings. This work provides a significant step forward in making gait recognition systems reliable for real-world applications