Graph-based Planning-informed Trajectory Prediction for Autonomous Driving

In the process of the autonomous driving, the accuracy of trajectory prediction of surrounding vehicles will significantly affect the downstream planning results of autonomous vehicles, and further affect the safety and efficiency of the traffic. Therefore, one of the major problems with autonomous driving is accurate trajectory prediction. However, the trajectory prediction task is challenging, mainly because the behavior of vehicles is influenced by many factors, such as the complexity of individual dynamics characteristics and the variability of spatial-temporal interactions between vehicles. In this paper, a planning-informed trajectory prediction method (GPiP) for autonomous driving is proposed to deal with the trajectory prediction problem of surrounding vehicles. More specially, the spatial-temporal graph convolutional network is proposed to encode the historical trajectories of all vehicles to extract the spatial-temporal features of the traffic graph. The planning coupled module is proposed to encode the future planning of autonomous vehicles to inform the trajectory prediction of surrounding vehicles. We evaluate our proposed method on NGSIM I–80 and US-101 datasets. The results show that our model is effective in trajectory prediction of surrounding vehicles of autonomous vehicle, and the integration of planning can improve the prediction accuracy.