Design And Implementation Of Trajectory Prediction System Based On Heterogeneous Graph Learning

Autonomous driving is one of the core technologies of the future intelligent transportation system.Due to the complex traffic scene,the accuracy of autonomous driving’s cognition of the surrounding environment will directly affect the effectiveness of autonomous driving decision-making.Therefore,how to accurately predict the trajectory of traffic participants in the traffic environment has become the core technical issue of autonomous driving.The complexity of the traffic scene is reflected in the complex interaction between various moving objects in the traffic environment,and with the different traffic scenes,the mutual influence between the objects is also different,which leads to the existence of many future motion trajectories of the objects in the traffic environment.Modal characteristics,it is difficult to accurately predict through the historical trajectory of a single object.In response to this problem,this paper proposes a heterogeneous graph learning trajectory prediction algorithm based on the distribution of destination points and the interaction of heterogeneous graph convolution features.Aiming at the multi-modal nature of future trajectories,this paper proposes a destination point distribution prediction method based on conditional variational encoder.By considering scene semantics and historical trajectory characteristics,it predicts the distribution of the end points of the trajectories of traffic participants,thereby introducing target constraints for trajectory prediction.Improve the performance and accuracy of trajectory prediction.Aiming at the trajectory uncertainty caused by the interaction between various types of moving objects,this paper designs a feature interaction model based on heterogeneous graph convolution.The model establishes the topological relationship between traffic participants as a heterogeneous graph,and models the interaction between moving objects through transformer-based information transfer and feature aggregation.Experiments show that the target point distribution prediction method and the feature interaction model based on heterogeneous graph convolution designed in this paper can effectively improve the trajectory prediction performance in complex traffic scenes.Finally,the paper designs and implements a trajectory prediction system based on vehicle-road coordination.Vehicle and roadside edge computing nodes perform local and global trajectory prediction respectively,and reduce prediction delay and improve computing efficiency through vehicle-road coordination.This paper analyzes the requirements of the trajectory prediction system based on heterogeneous graph learning,completes the general design and detailed design of the system,and displays the trajectory prediction results through the visual front-end,which verifies the effectiveness of the system.