Map Feature Based Trajectory Prediction with Multi-class Traffic Participants

Reliable prediction for the future trajectories of the surrounding multi-class traffic participants is critical for autonomous vehicles to drive safely and effectively in complex urban traffic. The existing methods based on Long-Short Term Memory (LSTM) networks mostly focus on the interactions between the same kind of traffic participants (i.e. homogeneous traffic participants), while ignoring the scene context and the interactions among different kinds of traffic participants (i.e. heterogeneous traffic participants). This paper proposes a novel method to take them all into account, modeling the interactions among heterogeneous traffic participants and utilizing the convolutional neural network (CNN) to extract map features. The results are evaluated on an autonomous driving dataset, Argoverse, and compared with two baseline methods. It is shown that our method has better performance and reduces the prediction error.