Low-Resource Spatial-Semantic Transformer Feature Extraction for Person Re-Identification in Crowded Video Surveillance
Annotatsiya
Person re-identification in crowded video surveillance is difficult due to occlusion, scale variation, viewpoint changes, and visually similar pedestrians. Existing deep models often provide strong accuracy but require high computational resources. This paper aims to develop a low-resource spatial-semantic feature extractor that produces compact and discriminative identity embeddings for real-time person re-identification. The proposed model processes each 3 × 192 × 64 full-body pedestrian crop using a 6 × 2 spatial grid and a semantic global branch. Local spatial tokens and global semantic features are fused, enhanced with positional encoding, and refined by lightweight transformer blocks. The final output is a 384-dimensional L2-normalized embedding trained using cosine-triplet learning. Experiments use a MOT20-derived dataset containing 387,426 pedestrian crops, 2,089 identities, and 310,602 triplets. The proposed model provides effective separation between same-identity and different-identity samples while maintaining compact representation and low computational cost. Similarity, distance, and threshold-based evaluations show that the spatial-semantic token design improves identity matching in crowded scenes. The proposed feature extractor offers an efficient solution for person re-identification in resource-constrained surveillance systems by combining local body-region information, global appearance features, and lightweight transformer refinement.
Hali tarjima qilinmagan