Rethink Reynolds' rules: flock-inspired network for vehicle trajectory prediction

Qifan Xue, Feng Yang, Shengyi Li, Xuanpeng Li, Guangyu Li, Weigong Zhang

Abstract: Trajectory prediction plays a pivotal role in the field of intelligent vehicles. However, due to the ignorance of future sequential and social correlations of vehicle trajectories, existing works usually suffer from serious inaccurate prediction, such as critical kinematic infeasibility and severe potential collision risks. To address the above challenges, this paper proposes an efficient flock-inspired network (FN)-based vehicle trajectory prediction algorithm. Specifically, first of all, by means of establishing the sequential correlations of vehicle trajectories, a feasible global velocity matching (GVM) scheme is proposed to enhance prediction efficiency. In addition, we design a viable radial basis attention (RBA) module to extract the active flock centering behaviors for improving the social compatibility of trajectory prediction. Moreover, in order to further reduce the potential risk of traffic collisions, we construct an available Hypothetical Rollout Predictor (HRP) to represent the passive avoidance behaviors among future trajectories. Finally, the extensive experimental results on four typical benchmark datasets (NGSIM, highD, INTERACTION, and Argoverse datasets) firmly demonstrate the effectiveness of our proposed algorithm compared with other state-of-the-art approaches.