Pedestrian trajectory prediction has a wide range of applications in areas such as autonomous driving and social robotics. Effective modeling of complex interactions between pedestrians is a key issue to improve trajectory prediction accuracy. However, when modeling the complex interactions between pedestrians based on graph neural networks, there are problems that the interactions between pedestrians do not change over time and the graph model cannot adjust the network parameters adaptively, resulting in large deviations between predicted and real trajectories. Therefore, this paper proposes a pedestrian trajectory prediction method based on dynamic evolving graphs, and designs dynamic feature update(DFU) to define the dynamic characteristics between pedestrians and model the dynamic interactions between pedestrians to build the network dynamics in the time domain, which improves the ability to model the complex interactions between pedestrians. An evolving graph convolution unit optimization encoder is used to flexibly evolve the graph model network parameters and enhance the adaptive capability of the graph model. The results show that the proposed model achieves better performance on two publicly available datasets(ETH and UCY) with 12.26% reduction in average displacement error and 14.10% reduction in final displacement error compared with the STGAT model at predicted 8 time steps.