차량 간 상호작용을 고려한 주변 차량 경로 예측

Abstract

In certain situations where lane changes take place frequently, such as at highway junctions or urban driveways, the dynamic interactions between neighboring and self-driving vehicles occur in a complex way. Therefore, it is necessary to consider both static elements such as lanes, and dynamic elements such as surrounding vehicles, for the trajectory planning of autonomous vehicles. This study suggests a deep learning network that predicts the trajectory of surrounding vehicles including ego-vehicles by considering the interactions between static and dynamic factors in the surroundings. The proposed network employed Encoder-Decorder structure, and applied additional Multi-head attention module, which is directly connected to the encoder, in effort to consider the interactions between self-driving and neighboring vehicles. Also, by applying masking technique, the situations where the number of vehicles observed and history are changing are taken into a consideration. Lastly, the proposed network has been verified through highD (Highway Drone) open dataset.