Abstract: Pedestrian re-identification searches for specific pedestrians in different environments, which has attracted widespread attention from domestic and foreign scholars in recent years. At present, pedestrian re-recognition algorithms mostly use a combination of local features and global features, and the training test on a single data set has achieved very good results. However, the results in the cross-domain test are not satisfactory, and the generalization ability is low. This paper proposes a cross- domain pedestrian re-recognition method based on deep capsule network. Through the view angle classification training task, the model can learn the effective features of the pedestrian in the image, and these features can be directly transferred to the pedestrian re-recognition task, alleviating the problem of insufficient generalization ability of pedestrian re-recognition. Experimental results show that this model is superior to all current pedestrian re-recognition methods based on unsupervised learning and has good generalization ability. 

Key words: pedestrian re-identification, cross-domain, visual angle, deep capsule network