Traffic data occupies an important position in intelligent transportation systems (ITS). However, the collected traffic data is often incomplete. We propose a generative adversarial network (GAN) model based on multi-perspective spatiotemporal learning (MST-GAN) to repair data. To achieve the effect of interpolating data from three perspectives: temporal, spatial, and spatiotemporal, we utilize chained generator with independent parameters to progressively refine the learning of temporal and spatial features. In addition, we achieve high-level fusion of multi-perspective features by adversarial between multiple generators and one discriminator. We conduct experiments on two real datasets, showing that the imputation effect of the MST-GAN model is better than other baseline models under different missing patterns. For example, the root mean square error (RMSE) is less than 7.5% and the mean absolute error (MAE) is less than 5% in the random missing scenario, which is much lower than the best performance error of other models.