How Will AVs Navigate Signalized Intersections?
Author: Robert Fischer, GTiMA President | Date: June 27, 2017
Autonomous vehicles promise such a degree of improved safety that we sometimes picture them as all-powerful, faultless machines. To a large extent, they will need to be. But how can we be sure they’re safe even in the most risky situations?
Consider the signalized intersection. Since their appearance at the end of the 19th century, traffic lights have been the primary mode of granting access to road intersections. However, traffic statistics show that, despite their claim to only a tiny percentage of road area, intersections are where 25% to 45% of all traffic collisions occur. Why is this?
When you think about it, it starts to become obvious. Intersections bring together cars traveling in all directions and then asks them to proceed through in groups timed by traffic signals. If any one car makes a false start, or speeds through the intersection and collides with a car in front of them, the entire exchange is compromised. And since cars are coming from all directions at various speeds, the risk of an accident that causes injury and/or major damage to vehicles goes way up, too.
So how will autonomous vehicles navigate the signalized intersection?
Therein lies the debate: will AVs independently “sense” their way through intersections using onboard sensors capable of discerning, for instance, the color of traffic lights, all while communicating with other vehicles (V2V) — or will AVs be “escorted” through intersections via connected infrastructure (V2I)?
As it turns out, the case for the V2V argument is in question. For starters, we would need to trust that every manufacturer has built in precise, latency-proof communications units that ensure constant communication with every other car. Statistically, the odds of that happening aren’t good, and it seems like too big of a risk when all it takes is the failure of one sensor on one car to cause an accident.
Furthermore, according to a paper by the Journal of Artificial Intelligence Research, “a vehicle approaching an intersection can quickly find itself in a situation where a collision is unavoidable, even when it has acted optimally.” In other words, even if no car overtly malfunctions, you can hardly call the intersection safe. As the JAIR paper notes, an intersection has very specific points of potential collision – hotspots that only one car at a time can occupy. That’s engineer-speak for “two cars occupying one hotspot will cause an accident.”
And this is why there’s a strong case for the V2I argument (where AVs get escorted through the intersection). The paper goes on to argue that an intersection needs its own “cooperative vehicle intersection control system (CVICS)” – hardware and software located at every interchange, which keeps an eye on all the points of potential collision and guides AVs through based on its safety-first algorithm.
Remember that acronym, CVICS, because it’s going to be a huge factor in the success of autonomous vehicles. It is absolutely essential for the safety of AVs, which can’t reasonably be expected to never run into trouble in a complex intersection, no matter how advanced they become.
Researchers at JAIR aren’t the only researchers arguing that CVIC systems are going to be a necessity for autonomous vehicles. There seems to be a consensus forming around the idea, to the point where researchers are now trying to determine the best algorithms for intersection control. MIT recently reported research on a “slot-based system”, similar to airplane-boarding procedures, designed for slow guidance through intersections in groups. They were essentially building on an IEEE research paper that offered a detailed algorithm for intersection control units to deploy.
The results from IEEE’s testing of their algorithm is worth a look. Not only was the flow through the intersection optimized for safety as expected, researchers also noticed some other improvements: vehicle stop delay time was reduced by an astonishing 99%, improving average travel time (under the presumption that a commute will involve several encounters with CVICS) by 33% or more. Carbon dioxide emissions were reduced by 44%, correlating directly with fuel savings of 44%. Basically, the researchers found that CVIC systems not only ensured safety, but made the process of getting through an intersection much faster. What’s not to like?
We remain optimistic about the safety benefits offered by autonomous vehicles. But the research is out there, and it’s compelling. Rather than risk public distrust in autonomous vehicles following a few crash-causing anomalies in intersections, it makes sense to deploy CVIC systems with proven algorithms to make intersections foolproof.
Just as we will eventually cede the wheel to our AVs, we will also need them to cede the wheel to CVICS from time to time and, in principle, the statistics showing such a disproportionate number of accidents occurring in intersections will shrink down to zero. We know that in the future our cars will be smart, but their interoperability with safety-oriented systems like CVICS will allow them to be brilliant.