Abstract: To address the complex modeling of dynamic interactions and the insufficient fusion of environmental semantic information in pedestrian trajectory prediction, a model named Environmental Semantics-aware Spatio-Temporal Graph Network (ES-STGNet) is proposed. This model constructs an end-to-end framework to deeply integrate dynamic social interactions among pedestrians with static physical constraints from the environment. Methodologically, a spatio-temporal graph is constructed for each scene in real-time. A Graph Convolutional Network is then utilized to explicitly capture the spatio-temporal dependencies among pedestrians. An information element spatial localization algorithm is also proposed. This algorithm dynamically integrates heterogeneous semantic information from high-definition maps, such as lane topology and traffic status, into the interaction encoder to enhance the physical plausibility of the predicted trajectories. The model employs a Laplacian Mixture Density Network as a decoder to generate multi-modal probabilistic trajectory predictions, effectively handling the uncertainty of pedestrian movement. Experimental results on the public ETH and UCY datasets show that the proposed model achieves superior performance on bothaverage displacement error(ADE) and final displacement error(FDE) metrics. This research confirms the effectiveness of combining environmental semantic information with a dynamic interaction model. It provides a more accurate and physically plausible trajectory prediction solution for autonomous driving and intelligent surveillance systems in complex urban scenarios.