As the most vulnerable traffic participants, pedestrians have always received considerable attention from autonomous driving. However, predicting the future behavior of pedestrians is challenging due to the intentions of pedestrian are potentially stochastic and difficult to be captured accurately through only a single trajectory. In order to solve these problems, we propose a multiple goals network (MGNet) for pedestrian trajectory prediction to generate a set of plausible trajectories in the crowds. The multimodality is achieved by sampling various goals from the parametric distribution which can sufficiently represent the stochastic intentions of pedestrian. The parametric distribution is obtained from the observations by a simple and effective multilayer perceptrons module. Finally, the whole future trajectories are generated by a Transformer-based encoder-decoder module with a new goal-visible masking mechanism. Experimental results on the most widely used datasets, i.e., the ETH-UCY datasets, demonstrate that MGNet is capable of achieving competitive performance compared with state-of-the-art methods.