Hybrid Proxy Aware Transformer Based Person Re-Identification Model for Student Surveillance

This work presents an improved person re-identification (Re-ID) model for student surveillance systems, based on a Hybrid Proxy Aware Transformer. We develop a methodology that extends conventional ViT outputs by deliberately rearranging and regrouping local tokens that are inspired by, using the Vision Transformer (ViT) as the fundamental architecture. Following their concatenation with class (CLS) tokens, these upgraded tokens become Proxy Aware Local Tokens (PALTs). These PALTs are then average pooled to yield a robust feature set that is optimized for individual Re-ID. We include Centroid Triplet Loss, which greatly enhances the feature embeddings’ discriminative ability, to further hone the model’s accuracy. Augments like random erasing evaluate the system’s ability to adapt to complicated monitoring settings and guarantee the model’s resilience to a range of environmental perturbations. To further enhance the model’s interpretability and dependability, we visualize significant regions using Gradient-weighted Class Activation Mapping (Grad-CAM). Initial findings reveal that our method performs better at re-identifying people in academic settings, demonstrating its potential for use in real-world surveillance scenarios.