Developing Decision-Making Models for AV Movements at the Unsignalized Intersections

Developing Decision-Making Models for AV Movements at the Unsignalized Intersections

James Sayer Headshot - link directs to their research page
James Sayer
Brian Lin Headshot - link directs their research page
Brian Lin
UMTRI Logo - link directs to U-M research page

Principal Investigator(s):

James Sayer, Director and Research Scientist – The University of Michigan Transportation Research Institute
Brian Lin, Assistant Research Scientist – The University of Michigan Transportation Research Institute

Project Abstract:
The goal of this research is to provide useful objective information for the development of autonomous vehicles (AV) and improve AV’s intelligence that can meet the other road users’ behavior and expectation in the unsignalized intersections. The research team mined UMTRI’s IVBSS naturalistic driving database to investigate drivers’ behavior and decisions when traversing the unsignalized intersections, which generally includes four-way and two-way stop intersections and T-intersection. The vehicle dynamic data and radar detection integrated with the video that showed the forward and side scenes formed a time-series variable set of three phases: just approach the stop bar, start to enter the intersection, and traverse through the intersection. A statistical decision-making method, Bayesian Network, and a machine-learning method, Mixture of Experts neural network, were used to model the maneuver (speed when entering the intersection, time/speed to traverse the intersection) and decision (waiting time at the stop bar) outputs for the traversal, with the time-series input data. By using Bayesian Network, the posterior mean and standard deviation of the outputs were estimated and could be updated when new prior input data was observed. With the neural network learning, the variables of the first phase were connected to those of the second and third phases. So far, the outcome was completely based on UMTRI’s naturalistic dataset and the structure of the networks was assumed. Therefore, the research team develop a modeling tool with MATLAB for the methods applied in this research. TRI can use the tool to customize networks by importing new variables, updating the network structure, training new models, and calculating posterior parameters with new evidences. It provides flexibility and potential to develop a smarter AV, as more and more on-road testing data is being collected. Besides traversing unsignalized intersections, other scenarios, such as zipper merging, left-turn across path, can be modeled once the data support. The final report for this project will not be publicly available.

Research Thrust(s): Control & Operations, Enabling TechnologyModeling & Implementation

Project Form(s):
Project Information Form