Pedestrian trajectory prediction is an important support for fully automatic driving in the city, and is widely used in robot path planning, autonomous cruise and other fields. The traffic scene from the driving perspective is complex and changeable, and the future position of pedestrians is uncertain. In this paper, a multi-information fusion network(MIFNet) is proposed to predict multiple possibilities of future pedestrian trajectories. Pedestrian posture information and optical flow information are used in the MIFNet on the basis of observed trajectory information, and the ways of reconstructing skeleton sequences and dividing local optical flow are used to avoid information distortion caused by pedestrian occlusion. In order to fuse these information more effectively, this paper proposes a cross- information fusion attention mechanism based on information evaluation. The importance of different information in the prediction process and the importance of different features between the same information are comprehensively considered. The MIFNet achieves the best results in predicting the average displacement error of 1.5 seconds on the PIE dataset, and the long-term trajectory prediction task of 1.5 seconds on the JAAD dataset. The prediction error is the smallest, and the number of model parameters and inference time are greatly reduced compared with the latest model.