부산대학교 도서관

Person re-identification plays an important role in the field of robot intelligence perception and public safety. Its main task is to identify person targets under cross-camera. However, the domain diversity between different datasets poses a clear challenge for adapting a model trained on one dataset to another. Currently, person re-identification methods based on domain adaptive learning and pseudo-label have made good progress on this problem. Unfortunately, inferior pseudo-labels and source domain noise affect the performance. In order to improve the quality of generated pseudo-labels and enhance the feature representation capability of the model, we propose a pseudo-label assisted optimization of a multi-branch person re-identification method. The multi-branch network is able to extract and represent more effective global and local features, and the generated pseudo-labels are optimized by using cosine similarity and DBSCAN clustering on the feature vectors, thus improving the consistency of the supervised information to enhance the cross-domain recognition performance. We also use a loss function combining cross-entropy loss and triplet loss to make the best feature learning. Experiments show that our method performs well in the Market-to-Duke and Duke-to-Market cross-domain recognition tasks.