<![CDATA[General Background: Gait recognition has emerged as a non-invasive biometric technique widely applied in forensic analysis, surveillance, and security systems. Specific Background: Recent advances in deep learning, particularly Generative Adversarial Networks and vision transformers, have enabled improved gait synthesis and feature representation, yet challenges remain in maintaining identity consistency and evaluation reliability. Knowledge Gap: Existing studies often address gait synthesis or recognition independently and lack unified frameworks combining generative quality assessment with robust feature extraction. Aims: This study proposes a comprehensive gait recognition framework integrating GAN-based gait enhancement, DINOv2 vision transformer feature extraction, and CNN-based identity verification. Results: Experimental evaluation on CASIA-B, OU-ISIR, and TUM-GAID datasets shows that StyleGAN2 achieves superior gait reconstruction quality, while DINOv2 provides highly discriminative spatio-temporal features, resulting in a classification accuracy of 98.3% and strong clustering separability. Novelty: The framework uniquely combines competitive GAN architectures with self-supervised transformer features and multi-level evaluation metrics within a single system. Implications: The proposed approach supports reliable gait recognition for biometric verification, forensic investigation, healthcare monitoring, and surveillance applications. Keywords: Gait Recognition, Generative Adversarial Networks, DINOv2, Vision Transformers, Biometric Verification Key Findings Highlights: High-fidelity gait synthesis was achieved using advanced generative architectures. Self-supervised transformer features provided strong identity separability. The integrated framework demonstrated robust performance across diverse conditions.]]>
