Graph-based trajectory prediction of surrounding vehicles using lane information

Abstract

In order to enable safe autonomous driving, it is necessary to predict the trajectory of surrounding vehicles. Only when accurately predicted trajectories of surrounding vehicles are utilized, autonomous vehicles can establish safe driving strategies and determine future behavior. In driving, driving behavior is affected by the road environment and interaction with surrounding vehicles. Therefore, it is difficult to accurately predict the trajectory without considering all of this information. In this study, we use graph neural network to predict the trajectory of surrounding vehicles, taking into account the interactions between surrounding vehicles. In addition, instead of using rasterized lane image information, we used vectorized lane information to lower the calculation cost and increase the accuracy of trajectory prediction. We also used graph neural network and attention to extract appropriate features from lane information. We created a dataset with data acquired in Sangam-dong, Mapo-gu, Seoul, and the proposed model was verified with this dataset. By configuring a recognition and tracking module, we demonstrated that the proposed model is robust to noise and can be operated in real-time in complex urban environment.