Abstract:

The prediction of peripheral target trajectories is an important basis for intelligent vehicle decision-making and planning. Existing modeling methods based on multi traffic agent Euclidean distance cannot effectively describe the complex interaction relationship between multiple targets， which limits the applicability in practical dynamic traffic scenarios. In this paper， complex network and memory-augmented neural network are innovatively integrated to construct a double-layer dynamic complex network model to achieve high reliability and interpretability of multimodal trajectory prediction. This model uses a Gaussian variable safety field to calculate risk weights， taking into consideration of the driving state parameters， shape and size of traffic participants， as well as the interaction between intelligent agents and the road， truly and accurately reflecting the interaction relationship between multiple traffic agents in complex environments. A complex network-coding module composed of attention mechanism and social pool containing risk weights is constructed to realize comprehensive and effective extraction of interaction features between traffic participants and road constraints in dynamic and complex scenes. A trajectory-decoding module based on reference trajectories is constructed， realizing multimodal trajectory output that balances accuracy and interpretability. The validation results on the public dataset nuScenes show that the method proposed in this paper has a minimum average displacement error of 1.37 m and a minimum final displacement error of 8.13 m， with excellent performance and good interpretability.

Key words: intelligent vehicles, trajectory prediction, complex networks, interactive modeling