关键词: 行人重识别, 域自适应, 多粒度特征挖掘, 域分布对齐

Abstract: In large-scale surveillance systems, the absence of positive cross-camera pedestrian samples in cross-regional scenes poses a limitation on the performance of person re-identification models. To tackle this challenge, an unsupervised domain adaptive person re-identification method incorporating multi-granularity feature mining and domain-invariant feature learning is proposed. The method comprises a multi-granularity feature learning module and a domain distribution alignment module. Within the multi-granularity feature learning module, global discriminant features of pedestrians are extracted through global features learning. To further enhance the discriminative features of pedestrians, a local consistency feature learning module is proposed to strengthen interactions among local features. Through the learning of both global and local features, the network is encouraged to extract multi-granularity discriminative features, thereby elevating the performance of the person re-identification model. Additionally, this study incorporates a domain distribution alignment module, conducting style transfer to construct positive samples with diverse styles across cameras for target domain. This not only addresses the issue of the lack of positive samples across cameras in cross-regional scenes but also enhances the domain adaptation capabilities of the model. Extensive experiments conducted on the Market-1501, DukeMTMC, CUHK03 and MSMT17 datasets demonstrate the effectiveness of the proposed method compared to state-of-the-art person re-identification methods.

Key words: person re-identification, domain adaptation, multi-granularity feature mining, domain distribution alignment