DGInet: Dynamic graph and interaction-aware convolutional network for vehicle trajectory prediction.

Area of Science:

• Computer Science
• Artificial Intelligence
• Robotics

Background:

• Existing graph convolutional network (GCN) models for trajectory prediction often lack temporal dynamics and complete interaction information.
• Accurately modeling spatio-temporal characteristics of dynamic graphs in traffic scenes remains a challenge.

Purpose of the Study:

• To propose a novel dynamic graph and interaction-aware neural network model (DGInet) for enhanced vehicle trajectory prediction.
• To address limitations in existing models by incorporating time attributes and complete interaction information.

Main Methods:

• Developed DGInet, a dual-network architecture combining a semi-global graph mechanism and an M-product based graph convolutional network.
• Utilized semi-global graph convolution for spatial interaction features and an M-product approach for dynamic graph extraction.
• Integrated semi-global network embeddings with M-product embeddings for final model representation.

Main Results:

• DGInet demonstrated superior performance compared to existing methods on the NGSIM and Apollo datasets.
• The proposed model achieved better prediction accuracy with reduced computational time.
• Effectiveness was further validated on the real-world Shenzhen traffic dataset.

Conclusions:

• DGInet effectively harnesses spatio-temporal dependencies and dynamic interactions for accurate vehicle trajectory prediction.
• The novel dual-network architecture and M-product approach offer significant improvements over conventional methods.
• DGInet presents a promising solution for real-world traffic prediction applications.