Interaction-Based Trajectory Prediction of Surrounding Vehicles with Driving Maneuvers Recognition

The accurate trajectory prediction of surrounding vehicles plays an important role for safe and comfortable driving of autonomous vehicles. An interaction-based trajectory prediction approach is proposed in this paper, which integrates online driving maneuvers recognition by using the self-attention mechanism combined with the feature extraction network. Specifically, the historical trajectories of surrounding vehicles are encoded by LSTMs, and then processed by a self-attention mechanism to obtain the correlation between them. The deep interaction features are extracted by the feature extraction network. The driving maneuver modes are defined as six basic types, and the predicted trajectory corresponding to the maneuver with the highest probability is the final result according to the extracted features. The maneuver prediction and the trajectory prediction results are evaluated on the NGSIM dataset and compared with other classical trajectory prediction models, the results clearly indicate that the proposed model can predict the driving maneuver with higher accuracy. The predicted trajectory and the real trajectory are visualized to represent the prediction results more intuitively and the predicted trajectory can well fit with the real trajectory.