<p>Occluded person re-identification (Re-ID) is severely hampered by incomplete visual cues from partial occlusions. This work proposes a novel framework, TRGA (Token Recovery and Global Augmentation), which learns robust representations via three synergistic components. The framework is founded on a Pedestrian Descriptor Extraction (PDE) module that distills robust descriptors without relying on fragile external models. PDE introduces a data-driven dynamic deletion strategy that analyzes an identity’s complete feature set and leverages intra-identity similarity to autonomously filter out noisy, occluded instances. This process yields robust descriptors adaptive to varying real-world occlusion ratios. These reliable descriptors then guide a Context-Aware Recovery Encoder (CRE), which employs a logical masking mechanism. In contrast to token pruning or expansion techniques that exclude tokens from computation, CRE retains the full token set throughout all stages, allowing all features to contribute to the global context as a progressive recovery process reveals informative tokens. Finally, an Identity-guided Global Augmentation (IGA) module performs a targeted, cross-identity augmentation, selecting the most confusing negative samples to simulate realistic inter-person occlusions. Extensive experiments on major occluded and holistic benchmarks validate TRGA’s superiority. On the Occluded-DukeMTMC dataset, the proposed framework establishes a new state-of-the-art, achieving 73.8% Rank-1 accuracy and 63.9% mAP.</p>