Quantitative Evaluation and Modeling of Action-Reaction Cycles in Interactive Human Driving Behaviors

This project aims to provide a framework to quantify the interactivity between human drivers in traffic, which could be used to design the decision-making policy for socially-compatible autonomous vehicles. Toward this end, we develop a conditional behavior prediction module based on Gaussian mixture models to jointly capture the dependency of one agent’s reactions on the other agent’s actions. We then implement an optimal transport theory to evaluate the interactiveness and dependency among traffic agents, resulting in a scalar value. Finally, we integrate the quantified dependency into an interaction-aware trajectory prediction model or controller to validate the performance in interactive traffic scenarios. This project would systematically and solidly bridge an enhanced understanding and characterization of human interaction-aware behaviors toward advances in autonomous car perception, decision-making, and control.

Faculty Supervisor:

Lijun Sun

Student:

Partner:

Carnegie Mellon University

Discipline:

Engineering

Sector:

Transportation (excluding aerospace); Automotive; Artificial Intelligence

University:

McGill University

Program: