Abstract: When vehicle re-identification models based on convolutional neural network perform convolutional and pooling operation, it is inevitable that the global sensitivity field will be narrow and local information will be lost. When illumination, perspective and resolution change dramatically, the robustness and accuracy of vehicle r re-identification will decline sharply. Therefore. the part coupled Transformer (PCT) network for vehicle re-identification is proposed, which stacks multiple PCT blocks to build a re-identification model. Each PCT block uses a part adaptation embedding (PAE) module to extract discriminative local features and employs a Transformer layer to extract robust global features. Firstly, the PAE that splits and adjusts dynamically feature map according to positions and scales, enhances the ability to capture local parts. Secondly, the self-attention mechanism of Transformer layers improves the representation of global features. Finally, the coupling between PAE module and Transformer layer results in an effective cooperation between global and local features. On VeRi-776 datasets and VehicleID datasets, CMC@1/CMC@5 of PCT reaches 0.970/0.988 and 0.865/0.985 respectively, which is superior to the comparison models.