Navigating intersections can be dangerous for humans and autonomous cars alike. In 2016, 23% of fatal and 32% of non-fatal incidents in the US occurred at intersections [1]. Our goal is to design new ways for a vehicle to understand the risk at an intersection, improving safety. Researchers Stephen McGill, Guy Rosman, and Luke Fletcher from Toyota Research Institute (TRI) collaborated with the MIT team of Teddy Ort, Alyssa Pierson, Igor Gilitschenski, Brandon Araki, Sertac Karaman, Daniela Rus, and John Leonard. Together, they created new state-of-the-art algorithms that better reason about the risk at intersections, and used the RACECAR to verify their algorithm experimentally.

Problem Setup

Our goal is to give the autonomous vehicle a better understanding of risk in traffic. We focus on intersections and left-hand turns, but we can also apply this understanding to roundabouts, merge lanes, and other tricky traffic situations.

For the autonomous car, we need to first define how it thinks about the road. As humans, we think about driving based on our current lane, and the future lane we’d like to drive in. We use a similar logic for the cars, and break the lanes into segments. The entire road becomes a series of segments, and the car checks that segments are empty before driving through them. At intersections, we observe all the lanes entering and exiting the intersection, and predict whether we think the intersection will be occupied or unoccupied when the car wants to enter.

The car also needs to understand what makes something risky. Here, our insight is we define risk as a combination of potential factors, then combine it all into a metric of risk. We consider: traffic density, physical occlusions, sensor noise, and the attention limitations of other drivers as possible sources of risk for the car. All of these risk factors can cause incidents for the car. For safe-driving, we want to minimize the chance of incidents.

Defining Risk

While we won’t go into the full details here, (please see the paper), the main contribution by McGill et al. is how to determine risk as the combination of risk over all segments and lanes. Overall, we consider the risk as the expected number of incidents the car would encounter if it drove along a part of the road.

Let’s define our intersection as having l ∈ {1,…, n_l } lanes, and within each of those lanes, we break that down into i ∈ {1, … , n_s } segments. Let E_{i,l}^t be an incident event within a segment, and let O_{i,l}^t be the observations the car makes using its sensors (cameras, lidar). The overall risk r_t is:

Here, the risk is the expected number of incidents, conditioned on the observations the car makes, over the segments and lanes. If the risk is low enough, then the car knows it is safe enough to drive through the intersection.

Note that we calculate the probability of an incident as a combination of our different factors: density, occlusions, noise, and attention. Please see the paper if you want more detail on how to model each of these factors.

Shared Control

Understanding the risk while driving doesn’t just help autonomous cars – it can help human drivers too! In our experiments, we also considered what it might look like if the car shares control with a human driver. In this case, we use the risk as a ``guardian angel’’ for the human driver. If the risk is too high, it keeps the human from making an unsafe decision (for example, preventing the car from entering the intersection if there’s too high of a risk of collision).

RACECAR Experiments

We tested our new algorithm in the real world using the RACECAR. In these experiments, we drive one ego car, but also use another six cars to generate cross-traffic, simulating a busy intersection. Overall, the team performed more than 100 trials of the car making an unprotected left turn across traffic. In the video below, we see how the car can use the risk to better navigate this tricky intersection, both when sharing control with a human driver and when fully autonomous.

[1] Incidents totaled 6,462 fatal and 10,119 non-fatal crashes, per the National Highway Traffic Safety Administration, “Fatality Analysis Reporting System (FARS) Encyclopedia,” Available: https://www-fars.nhtsa.dot.gov

Paper link: https://ieeexplore.ieee.org/abstract/document/8779655