Abstract: Vehicle re-identification (ReID), as a critical component for intelligent transportation systems, aims to identify the same vehicle across different cameras with large variations in viewpoints, lighting, and resolution. As the rapid development of surveillance networks, a tremendous amount of data is collected, making it expensive or even impossible to retrieve the same vehicle across multiple cameras via human labor. With the breakthrough in computer vision, a variety of vehicle ReID methods based on convolutional neural networks are recently developed and perform well in real-world scenes. One of the main challenges of ReID task is the variation of viewpoints, under which circumstance the identical vehicle from different viewpoints usually has low intra-instance similarity while different vehicles from similar viewpoints may have relatively small inter-instance discrepancy. To address this challenge and improve vehicle ReID performance, a novel parsing-based vehicle part multi-attention adaptive fusion network (PPAAF) is proposed. First, in order to gain a fine-grained representation of vehicle parts and meanwhile eliminate the effects of viewpoints, we modify the existing deep residual network named ResNet and introduce a parsing network to segment a vehicle image into image background and four different vehicle parts (including front, back, top, and side) respectively. Then vehicle part features are aligned through masked average pooling. Second, a novel part attention module (PAM) is developed to integrate multiple attention mechanism adaptively, providing diverse measurements of vehicle part importance in vehicle ReID task. Compared with previous methods, the additional attention not only enlarges the distance between different vehicles but also strengthens the validity of vehicle part feature extraction. Finally, we optimize the proposed network with Triplet loss and Focal loss in multi-task framework. Extensive experiments and ablation studies conducted on two major vehicle re-identification datasets provide competitive results against state-of-the-art vehicle ReID methods and prove the effectiveness of our method.