Graph Neural Networks for Modelling Traffic Participant Interaction

Abstract

By interpreting a traffic scene as a graph of interacting vehicles, we gain a flexible abstract representation which allows us to apply Graph Neural Network (GNN) models for traffic prediction. These naturally take interaction between traffic participants into account while being computationally efficient and providing large model capacity. We evaluate two state-of-the art GNN architectures and introduce several adaptations for our specific scenario. We show that prediction error in scenarios with much interaction decreases by 30% compared to a model that does not take interactions into account. This suggests that interaction is important, and shows that we can model it using graphs. This makes GNNs a worthwhile addition to traffic prediction systems. †

Deep Learning-Based Vehicle Behavior Prediction for Autonomous Driving Applications: A Review

“…For example, in a traffic jam, knowing that the second vehicle ahead of the TV is accelerating can enable early prediction of speed increase for the TV. Instead of considering a fixed number of vehicles as the SVs, a distance threshold is defined to divide vehicle into the SVs and NVs in [29], [30], [31]. It means that only the interactions of vehicles within this threshold are considered in the prediction model.…”

“…[28] History of states for the TV and nine SVs. [29], [30], [31] A distance threshold is defined to divide vehicles into the SVs and NVs. [32] A soft attention mechanism is used to weight the impact of each observed vehicle.…”

“…The straightforward approach to predict the trajectory of the TV is to estimate the position of it over time [47], [38], [29]. The predictor model can also estimate the displacement of the TV relative to its last position at each step [31], [25]. The other approach used in [28] is to predict lateral position and longitudinal velocity separately.…”

“…Using this representation, Graph Neural Networks (GNNs) [65], [66] can be used to predict TV's behaviour. Diehl et al [31] compare the trajectory prediction performance of two state-of-the-art graph neural networks, namely, Graph Convolutional Network(GCN) [50] and Graph Attention Network (GAT) [51]. They also propose some adaptations to improve the performance of these networks for the vehicle behaviour prediction problem.…”

Dynamic Input for Deep Reinforcement Learning in Autonomous Driving

“…Edge weights are computed by the inverse absolute distance between two vehicles, as shown in [23]. A fullyconnected graph is avoided due to computational complexity.…”