Abstract:Accurate prediction of vehicle trajectories is crucial to the safety and comfort of autonomous vehicles. Although several graph-based models have exhibited substantial progress in acquiring spatiotemporal dependencies among vehicles in the driving environment, the potential for additional exploration in this domain persists. The main reason is that they concentrated on independently capturing the spatial relations and temporal dependencies, neglecting to incorporate the temporal feature into the spatial feature for co-training, which limits their ability to yield satisfactory predictive accuracy. Typically, spatial and temporal correlations are coupled and should be modeled jointly. Inspired by this, a novel dynamic graph neural network with spatial–temporal synchronization (STS-DGNN) for vehicle trajectory prediction is proposed, which constructs the driving scene as dynamic graphs and can jointly extract spatial–temporal features. Specifically, low-order and high-order dynamics of vehicle trajectories are considered collaboratively in a one-stage framework rather than independently modeling the spatial relationship and temporal correlations of vehicles in two-stage models. The proposed model also considers the dynamic nature of graph sequence by utilizing gate recurrent unit (GRU) to update the graph neural network (GNN) parameters dynamically. The spatial–temporal features are subsequently conveyed to convolutional neural networks (CNNs) and processed by a multilayer perceptron (MLP) to generate the ultimate trajectories. Finally, to illustrate the effectiveness of the STS-DGNN model, the model is assessed on three well-known datasets, namely highD, EWAP, and UCY. The results confirm that our model performs better at making predictions than cutting-edge models. The visualization results intuitively explain that our method can extract sophisticated and subtle multivehicle interactions, resulting in accurate- predictions.

Understanding Graph Neural Networks via Trajectory Analysis

How Graph Neural Networks Learn: Lessons from Training Dynamics

GRLSTM: Trajectory Similarity Computation with Graph-Based Residual LSTM

Graph Neural Networks Inspired by Classical Iterative Algorithms

Unsupervised Pedestrian Trajectory Prediction with Graph Neural Networks.

Toward the Analysis of Graph Neural Networks

Explaining GNN over Evolving Graphs using Information Flow

Revisiting graph neural networks from hybrid regularized graph signal reconstruction

Neural Trees for Learning on Graphs

Learning to Identify Graphs from Node Trajectories in Multi-Robot Networks

GNN-RM: A Trajectory Completion Algorithm Based on Graph Neural Networks and Regeneration Modules

How Neural Networks Extrapolate: From Feedforward to Graph Neural Networks

Graph Neural Networks: Architectures, Stability, and Transferability

Trajectory Planning for Autonomous Driving in Unstructured Scenarios Based on Graph Neural Network and Numerical Optimization

Multi-Scene Generalized Trajectory Global Graph Solver with Composite Nodes for Multiple Object Tracking

Foundations and Frontiers of Graph Learning Theory

Towards Understanding Graph Neural Networks: An Algorithm Unrolling Perspective

STS-DGNN: Vehicle Trajectory Prediction via Dynamic Graph Neural Network With Spatial–Temporal Synchronization

Training Graph Neural Networks on Growing Stochastic Graphs

Two-Stage Training of Graph Neural Networks for Graph Classification

Spatio-Temporal Graph Convolutional Networks via View Fusion for Trajectory Data Analytics