How Many Lives Could Self-Driving Cars Save? This Tool Predicts It

Promoters of self-driving technology say it could prevent most of the traffic deaths that happen today, often citing the fact that 94% of all crashes in the US are caused by human error. But will all those deaths ever be prevented? And what’s the best way to get there?

To make predictions as realistic as possible, regulators and the industry should look at a wide range of factors, according to researchers at the University of Kentucky and the Kentucky Transportation Center. These two organizations are developing a tool called Data-Driven Safety Assessment for Connected and Autonomous Transportation (DDSAFCAT) that could bring about a dozen variables into play in answering these questions. Regulators and others could use it to test different assumptions and reach conclusions that could inform choices about how much automation to allow, encourage or implement.

The issue of how and when automation will make vehicles safer returned to the spotlight this week when the Insurance Institute for Highway Safety revealed test results that showed semi-autonomous features in some production cars, such as automatic lane-keeping and adaptive cruise control, often failed to respond correctly to obstacles such as lane markers and stopped cars. IIHS concluded the features weren’t good substitutes for human drivers and in some cases could be dangerous.

How many fatalities self-driving or semi-autonomous cars can prevent will depend on which features are introduced, how human drivers respond to them, how well those features can deal with what really causes crashes, well as other variables, according to the Kentucky researchers. They revealed their work on a prediction tool in a poster presented at the Automated Vehicles Symposium in San Francisco last month.

High-level predictions about the benefits of automation have helped to generate support for technology that is likely to save lives, but they have limitations, said Austin Obenauf, a graduate student at Kentucky’s school of civil engineering who is working on the project along with Professor Reginald Souleyrette and graduate student Freddy Lause.

For example, the 94% of deaths caused by human drivers is probably in the upper boundary of what self-driving cars could prevent, because the way they outperform human drivers may not always be enough, he said. If a pedestrian walks into the street, an AV might detect the person more quickly than a human driver would but still not be able to stop in time due to the laws of physics, Obenauf said.

The Kentucky team’s tool is currently implemented in Microsoft Excel using a logistics function to create graphs based on inputs. It would allow safety officials or anyone else to combine historical data from sources such as accident statistics with assumptions about the future. By changing those assumptions, they could come up with different predictions that could inform policy decisions. The research company Rand introduced a similar tool last year, but this one would allow for more variables, Obenauf said.

The researchers are building the tool to predict results over 15 years, but users could apply longer timeframes if, for example, they expected slower adoption of autonomous vehicles.

Some public safety agencies have rich collections of data about where and how accidents have happened, which can help in predicting the effect of automation, Obenauf said.

For example, if the data showed there were more rear-end collisions than sideswipe crashes, then it might be possible to predict that emergency braking features would prevent more accidents than lane-keeping features.

At a higher level, the tool might predict that having advanced driver assistance systems in more human-driven cars would prevent a certain number of crashes over the next five years. If a user assumed that fully autonomous cars would prevent even more crashes, but wouldn’t be widely available for more than five years, then they might conclude that more lives could be saved by encouraging driver assistance features rather than waiting for full AVs.

Users could also try out different assumptions about future variables, such as how effective they think AVs will be at preventing crashes, how quickly consumers would adopt automation, and how many would drive less carefully if they perceived less risk due to driver assistance features. As the reality of self-driving cars plays out, what were once future assumptions will become historical data that can’t be changed but continues to shape conclusions about the future.

— Stephen Lawson is a freelance writer based in San Francisco. Follow him on Twitter @sdlawsonmedia.

Autonomous Vehicles