SELF-SUPERVISED GAIT RECOGNITION WITH DIFFUSION MODEL PRETRAINING

Gait recognition has potential as a biometric tool but faces challenges in unconstrained environments due to occlusions and varying conditions. We propose a novel approach using diffusion models for self-supervised pretraining to enhance gait recognition accuracy. The method involves pretraining a gait feature extractor with a conditional latent diffusion model, followed by ne-tuning for specic tasks. Experiments on Gait3D and GREW datasets show signicant performance improvements, demonstrating that diffusion model pretraining offers better feature representation and recognition in real-world scenarios. This approach opens new avenues for robust gait analysis.