Roundabout Connected/Automated Vehicle Active Inference Control Strategies to Improve Safety for All Users

Roundabout Connected/Automated Vehicle Active Inference Control Strategies to Improve Safety for All Users

Headshot of Vassilios Morellas. The link directs to their profile page.
Vassilios Morellas
Headshot of Ted Morris. The link directs to their profile page.
Ted Morris
University of Minnesota Logo. The link directs to the funded research led by this institution.

Principal Investigator(s):

Vassilios Morellas, Research Professor of Electrical and Computer Engineering – University of Minnesota
Ted Morris, Senior Engineer in Computer Science & Engineering – University of Minnesota

Project Abstract:
Many innovations have recently been adopted for new and existing urban roadways that improve safety for all users, with concomitant goals to also reduce congestion. Roundabout intersections, which have been adopted in many countries for decades, are becoming one such innovation adopted in the US. Voluminous studies confirm roundabouts reduce dangerous traffic conflicts between human-operated vehicles, and their speeds, thereby potentially reducing serious crashes. One can appreciate the perceptual complexities presented to a human driver within these intersections when confronted with appropriate gap selection decisions simultaneously with vulnerable road user (VRU) interactions (varied behaviors of approaching bicyclists, public transit, and pedestrian crossings, for example). Automated vehicles in general will need to embed such perception-action behaviors at these intersections to correctly react to VRUs as well as likely interactions with human-driven vehicles several years into the future.

We propose to develop Active Inference Connected/Autonomous Vehicle (CAV) control strategies to reduce speed according to anticipated vehicle and pedestrian actions using real-time roadside sensor observation data. Originally grounded in neuropsychology and physiology, Active Inference is a probabilistic framework that contends perception, learning, and decision-making (and the resulting actions) are interdependent forms of inference. An agent (CAV) infers future actions most likely to generate preferred observations (states of all users and itself) concomitantly with sequences of actions that balance reducing uncertainty while encouraging learning. The mathematical framework will require significant observational data to formulate and validate the learned perception and decision models, as well as address computational challenges for the vehicle and edge processing. Accordingly, a two-phase study is proposed to address this problem. The first phase in year 1 deploys and evaluates roadside sensing (LiDAR and camera sensors) to accurately detect, edge-process, and package estimates of all user states in order to broadcast them through generated Basic Safety Messages (BSM). The roadside sensing challenge is to provide reliable ‘eyes’ to where the vehicle cannot adequately ‘see’ due to line of site limitations. The BSMs can alert human drivers to potential conflicts, such as far-side pedestrian crossing events, for example, that may not be as visually evident to human drivers. A second research phase will then focus on the automated vehicle control strategies (i.e., reduce its speed, and invoke yield decisions accordingly), using the complete road user traffic states provided by the roadside detection. The AI algorithm will be developed, tested, and demonstrated with the U of MN C/A research vehicle.

Institution(s): University of Minnesota

Award Year: 2023

Research Focus: Safety, Mobility, Equity

Project Form(s):