Truly autonomous vehicles — those without human safety drivers at the wheel — must be capable of determining when it’s safe to merge into traffic. Intersections with obstructed views make this somewhat challenging, but researchers at Toyota and MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL) say they’ve developed an AI model that can estimate collision risk highly accurately.
The key turned out to be uncertainty. The model — which the team designed specifically for junctions that lack a stoplight, forcing cars to yield for traffic — weighs several factors in tabulating risk, including visual occlusions, sensor noise and errors, the speed of other cars, and the attentiveness of other drivers. It also considers how long it’d take the car to steer through the intersection, along with all safe stopping spots for crossing traffic.
The model splits the road into segments, enabling it to determine if any one section is occupied by another car. The moment a passing car travels into a segment, its speed informs a prediction of the car’s progression through subsequent segments. Simultaneously, the model considers the road segments the car passed through before the intersection, the rationale being that cars occupying a high number likely spotted the autonomous car.
The aforementioned risk estimate is updated continuously. In the presence of multiple occlusions, the model directs the car to nudge forward in order to reduce uncertainty. And when the risk bottoms out, the model has it drive through the intersection without stopping so as to avoid increasing the risk of collision by lingering.
The team managed to run the model on remote-control cars in real time, suggesting it’s efficient and fast enough to deploy into full-scale autonomous test cars in the near future. They concede that it needs more rigorous testing, but they believe it could serve as a supplemental risk metric that an autonomous vehicle system can use to better reason about driving through intersections safely.
“When you approach an intersection there is potential danger for collision. Cameras and other sensors require line of sight. If there are occlusions, they don’t have enough visibility to assess whether it’s likely that something is coming,” said director of CSAIL Daniela Rus in a statement. “In this work, we use a predictive-control model that’s more robust to uncertainty, to help vehicles safely navigate these challenging road situations.”
As a next step, the researchers aim to incorporate risk factors such as the presence of pedestrians in and around the road junction in the model.