3D-MBNet: Intention Based Multimodal Vehicle Trajectory Prediction with 3D Social Convolution

Predicting vehicle trajectories in traffic scenes is an important issue for autonomous driving and advanced driver assistance systems. The spatial and temporal interactions between the predicted vehicle and its surrounding vehicles are commonly used for trajectory prediction. However, most existing methods usually deal with the two interactions separately. In this paper, we propose a multimodal vehicle trajectory prediction model, which includes a 3D social convolution module to jointly model spatial and temporal interactions. Furthermore, to make interpretable multimodal predictions, we define several types of driving intentions (namely lane changes, acceleration, deceleration and driving with uniform velocity) and introduce multi-branch decoders to decode different intentions. The two-stage training strategy is introduced to guarantee the recall rates of the intentions with long-tailed distributions. Our model outperforms several state-of-the-art methods on the NGSIM and highD benchmark datasets. In addition, the ablation experiments show that our method can improve the prediction accuracy of each specified modality.