MMTP: Multi-Modal Trajectory Prediction with Interaction Attention and Adaptive Task Weighting

Accurate prediction of the driving intentions and trajectories of other vehicles is critical to the planning and control subsystem of the autonomous driving system. In addition to the driver's driving habits, the future driving intention and trajectory of a vehicle are the result of dynamic interactions with others around it, and the driver should have multiple executable driving trajectories to choose from at a given moment. In this paper, we propose a new interaction attention mechanism and a lightweight multi-modal maneuver-based trajectory prediction model. In addition, we consider it as multi-task model and put forward an adaptive task loss weighting scheme for further performance improving. We evaluate our method on dataset NGSIM US101, and the results show that the proposed model achieves the optimal performance, the lowest model complexity and our task loss weighting scheme can further improve the model performance compared to the original task loss scheme.