Abstract: Cluster-based unsupervised person re-identification generates pseudo labels through clustering and constructs a single proxy for each cluster. However, inter-camera discrepancies between different cameras and intra-camera differences caused by pedestrians' different stepping directions under the same camera make it impossible for a single proxy to adequately represent the whole complex cluster, affecting the training process and model performance. Thus, a camera-aware diverse proxies based unsupervised person re-identification method is proposed to eliminate inter-camera discrepancies and intra-camera differences in order to improve the intra-class tightness. Camera-aware diverse proxies include camera multi proxies and camera discrepancy proxy. The camera multi proxies construct a mean proxy and a hard proxy for each camera class to improve intra-camera tightness through complementary representation. A camera discrepancy proxy is also constructed for each camera class, which utilizes the mean proxies of different camera classes for updating, so as to bring cameras closer to each other and reduce camera discrepancy. The camera discrepancy proxy and camera multi proxies collaborate with each other to improve intra-class tightness and model performance. The proposed method achieves 90.4%/96.4%, 75.8%/86.5%, and 58.2%/82.5% for mAP/Rank-1 on the Martket1501, DukeMTMC-reID, and MSMT17 datasets, respectively.