그래프 어텐션 기반 에이전트 상호작용을 고려한 경로 예측

Abstract

Considering nearby agents such as vehicles, pedestrians and motorcycles is essential in path planning of self-driving cars. Agents do not move independently, so path predictions reflecting interactions are needed. In this paper, we propose a network to predict the paths of traffic agents considering interactions. It only considers close agents based on distance among thesurrounding agents to predict the path. In the proposed model, agent information of each frame is expressed as a graph and given as a input to a Graph Attention Network to generate interaction-embedded agent information in each frame. Interactions with highly relevant agents can be considered by applying attention in the network to the graphs. This information goes into the input of the Sequence-to-Sequence structured encoder-decoder and predicts the future path of each agent simultaneously. Lastly, the proposed network has been verified through the Apolloscape and HighD open dataset.