In order to effectively solve the relocalization problem of mobile robot, a robot relocalization method fusing deep learning and particle filtering is proposed. Firstly, a 3-DOF mobile robot relocalization framework is proposed, which mainly includes two progressive stages: relocalization model construction and robot online relocalization. Secondly, a 3-DOF mobile robot relocalization network model, G_PoseNet, is proposed and constructed based on PoseNet. The pose result predicted by G_PoseNet is used as the initialization state of particle filter localization algorithm to support the subsequent relocalization process. Then, a data model based kidnapping state judging method is proposed to determine whether to start the relocalization process. Finally, a large number of experiments on public datasets and real environment were performed to verify the proposed method. The result shows that the G_PoseNet model can guarantee a certain degree of location prediction accuracy and improve the pose angle prediction accuracy, the success rate of robot relocalization achieves 87% .