[Paper Review] A Taxonomy and Review of Algorithms for Modeling and Predicting Human Driver Behavior

Bhattacharyya, Raunak P., Kyle Brown+3 more|arXiv (Cornell University)|Jun 15, 2020

Autonomous Vehicle Technology and Safety30 citations

TL;DR

This paper presents a comprehensive taxonomy and review of 200 driver-behavior models within a unified POSG-based framework, detailing core tasks (state estimation, intention estimation, trait estimation, and motion prediction) and auxiliary tasks.

ABSTRACT

An open problem in autonomous driving research is modeling human driving behavior, which is needed for the planning component of the autonomy stack, safety validation through traffic simulation, and causal inference for generating explanations for autonomous driving. Modeling human driving behavior is challenging because it is stochastic, high-dimensional, and involves interaction between multiple agents. This problem has been studied in various communities with a vast body of literature. Existing reviews have generally focused on one aspect: motion prediction. In this article, we present a unification of the literature that covers intent estimation, trait estimation, and motion prediction. This unification is enabled by modeling multi-agent driving as a partially observable stochastic game, which allows us to cast driver modeling tasks as inference problems. We classify driver models into a taxonomy based on the specific tasks they address and the key attributes of their approach. Finally, we identify open research opportunities in the field of driver modeling.

Motivation & Objective

• Establish a common mathematical framework for modeling driver behavior in multi-agent traffic.
• Classify existing driver models using a taxonomy centered on core tasks and auxiliary capabilities.
• Map 200 published models into the taxonomy to illuminate connections and gaps for researchers.

Proposed method

• Introduce a discrete-time multi-agent partially observable stochastic game (POSG) framework to describe interactive traffic dynamics.
• Define core modeling tasks: state estimation, intention estimation, trait estimation, and motion prediction, with explicit inference targets.
• Provide task-specific comparison tables and keyword catalogs to position models along architecture, training, theory, scope, and evaluation axes.
• Differentiate between online/offline trait estimation and between motion prediction (online, discriminative) and traffic simulation (offline, generative).
• Summarize existing models by their approach to observation, internal state, policy, and state-transition assumptions.

Experimental results

Research questions

• RQ1 How can driver behavior modeling in multi-agent traffic be unified under a common POSG-based mathematical framework?
• RQ2 How do existing models differ across core tasks (state estimation, intention/trait estimation, motion prediction) and auxiliary tasks, and where do they fit within a structured taxonomy?
• RQ3 What are the tradeoffs between different intention spaces, hypothesis representations, and inference paradigms across driver models?
• RQ4 What guidance does the taxonomy provide for selecting models for state estimation, intention estimation, trait estimation, and motion prediction in automated-vehicle planning and control?

Key findings

• A POSG-based formulation provides a unifying mathematical framework for describing microscopic, interactive traffic dynamics.
• A taxonomy is constructed around four core tasks (state estimation, intention estimation, trait estimation, motion prediction) plus auxiliary tasks (risk estimation, anomaly detection, behavior imitation, microscopic simulation).
• The paper catalogues and categorizes 200 driver models, detailing their architecture, training, theory, scope, and evaluation characteristics.
• Intention estimation models vary in their intention spaces, hypothesis representations, and inference paradigms, including Bayesian, prototype-maneuver, and game-theoretic approaches.
• Trait estimation is often offline or online and can be represented by deterministic parameters, with some models employing particle or continuous distributions for parameter uncertainty.