[Paper Review] Modeling Multi-Vehicle Interaction Scenarios Using Gaussian Random Field

Yaohui Guo, Vinay Varma Kalidindi+5 more|arXiv (Cornell University)|Jun 25, 2019

Autonomous Vehicle Technology and Safety14 references4 citations

TL;DR

This paper proposes a non-parametric Bayesian framework that models multi-vehicle interaction scenarios using a Gaussian Process (GP) mixture with a Dirichlet Process (DP) prior to learn motion patterns from naturalistic traffic data. The method captures complex, high-dimensional driving interactions without pre-specifying scenario types, enabling realistic simulation of highway and intersection scenes with implicit road layout learning.

ABSTRACT

Autonomous vehicles are expected to navigate in complex traffic scenarios with multiple surrounding vehicles. The correlations between road users vary over time, the degree of which, in theory, could be infinitely large, thus posing a great challenge in modeling and predicting the driving environment. In this paper, we propose a method to model multi-vehicle interactions using a stochastic vector field model and apply non-parametric Bayesian learning to extract the underlying motion patterns from a large quantity of naturalistic traffic data. We then use this model to reproduce the high-dimensional driving scenarios in a finitely tractable form. We use a Gaussian process to model multi-vehicle motion, and a Dirichlet process to assign each observation to a specific scenario. We verify the effectiveness of the proposed method on highway and intersection datasets from the NGSIM project, in which complex multi-vehicle interactions are prevalent. The results show that the proposed method can capture motion patterns from both settings, without imposing heroic prior, and hence demonstrate the potential application for a wide array of traffic situations. The proposed modeling method could enable simulation platforms and other testing methods designed for autonomous vehicle evaluation, to easily model and generate traffic scenarios emulating large scale driving data.

Motivation & Objective

• To develop a fully data-driven method for modeling complex, high-dimensional multi-vehicle interactions in autonomous vehicle environments.
• To overcome limitations of traditional models that rely on strong priors, fixed scenario counts, or simplified one-on-one interactions.
• To enable the extraction of underlying motion patterns from large-scale naturalistic traffic data without assuming the number of interaction scenarios in advance.
• To generate realistic, representative traffic scenarios for autonomous vehicle testing and simulation.
• To implicitly learn road layout and lane-level behavior from trajectory data without explicit geometric input.

Proposed method

• Model multi-vehicle motion as a mixture of Gaussian Process (GP) components, where each component represents a distinct motion pattern.
• Use a Dirichlet Process (DP) prior to non-parametrically infer the number of motion patterns from data, avoiding pre-specification.
• Apply Gibbs sampling for posterior inference over mixture components, weights, and GP hyperparameters.
• Represent each motion pattern as a GP velocity field over the region of interest (ROI), enabling smooth, continuous trajectory generation.
• Train the model on NGSIM highway and intersection datasets to learn interaction dynamics from real-world observations.
• Use the learned GP mean velocity fields to simulate trajectories that reflect observed interaction scenarios, including lane usage and turning behaviors.

Experimental results

Research questions

• RQ1 Can a non-parametric Bayesian model effectively learn multi-vehicle interaction patterns from large-scale naturalistic traffic data without assuming a fixed number of scenarios?
• RQ2 How well can a Gaussian Process-based velocity field model capture complex, multi-directional interactions at intersections and on highways?
• RQ3 To what extent can the model implicitly learn road layout and lane-level behavior from trajectory data alone?
• RQ4 Can the learned motion patterns generate realistic, diverse traffic scenarios that reflect real-world driving behavior?
• RQ5 How does the model perform in capturing dynamic, high-dimensional interactions compared to parametric or simplified models?

Key findings

• The model successfully learned motion patterns from both highway and intersection datasets, capturing diverse interaction types such as merging, lane changes, and left turns.
• The GP mean velocity fields implicitly learned road boundaries and lane configurations without explicit geometric input, as evidenced by low probability mass outside road areas.
• Simulated trajectories closely matched real-world behavior, including directional flow and stop-and-go patterns at intersections.
• The model generated motion patterns that reflected semantic traffic scenarios—e.g., vehicles turning left, going straight, or waiting at intersections—without manual labeling.
• The Dirichlet Process enabled automatic discovery of the number of underlying motion patterns, avoiding the need for prior assumptions about scenario count.
• The framework demonstrated robustness in modeling complex, multi-directional interactions, particularly in intersection scenarios with higher interaction diversity than highway data.