Abstract:

Accurate prediction of the future trajectory of surrounding vehicles is crucial to the decision-making and motion planning of autonomous vehicle. Existing research tends to use Recurrent Neural Networks （RNN） to model the time interaction of vehicles， but its interpretability of vehicle interaction modeling is poor， ignoring the actual lane structure， and there are deficiencies in capturing the interaction between vehicles and the environment. To address this problem， in this paper， a vehicle trajectory prediction model based on graph convolutional interactive networks that considers lane topology constraints is proposed. The vehicle interaction relationship extraction module adds edge weights when constructing the spatial relationship of vehicles to consider their neighboring interaction， making the interaction more interpretable. The driving scene representation module aims to improve the accuracy of vehicle trajectory prediction by extracting lane topology from high-precision maps. The trajectory prediction module integrates the output of the above two modules and outputs the predicted future trajectory. This integration allows for more precise modeling of the interaction between road structures and vehicle driving trajectories. The experimental results show that compared with mainstream methods， this model has achieved good performance on the Argoverse dataset， improving the accuracy and rationality of vehicle trajectory planning under complex road structures.

Key words: autonomous vehicle, trajectory prediction, graph convolution network, interaction behavior