UNSUPERVISED PERSON RE-IDENTIFICATION BASED ON LOCAL FEATURE ENHANCEMENT AND CROSS-VIEW NEIGHBOR CLUSTERING

Existing unsupervised person re-identification methods ignore the pedestrian occlusion, pose variants and different importance between local features. To solve these issues, this paper proposes an unsupervised local feature enhancement and cross-view neighbor clustering network LECNN. The LECNN comprehensively modeled pedestrians, local and global features based on the spatial attention mechanism, strengthened more discriminative local features, and obtained a more representative comprehensive feature representation. In addition, the proposed method constructed a cross-view neighbor matrix to calculate cross-view positive and negative samples of pedestrians, achieving more accurate and robust clustering results. Compared with the state of arts methods, the proposed method achieves higher person recognition rates in 2 datasets, and effectively reduces the effects of pose variants and pedestrian occlusion.