Estimating drivers’ behaviors from event recorder data
You could call it CSI: Crash Scene Investigation—a new project by U of M researchers sits at the intersection of crash investigation and forensic science.
For more than a decade, the transportation community has recognized that using pre-crash data from event data recorders (EDRs) installed in many passenger vehicles might help illuminate the actions of drivers prior to crashes. Now, that recognition is becoming reality.
“In forensic science, we need to quantify uncertainty,” says Gary Davis, a professor in the Department of Civil, Environmental, and Geo- Engineering. “Often, the crime is known but the criminal is unknown. However, when investigating car crashes we have the opposite problem—we know who was driving the cars; the question is whether there was a crime.”
Using data from left-turning crashes where pre-crash data are available from both vehicles, researchers set out to determine whether they could estimate features such as the location and speed of the opposing vehicle at the time of turn initiation and the reaction time of the opposing driver. To do this, they needed to overcome a number of difficulties, including data measurement errors and the fact that EDR data from the two vehicles are not synchronized.
To help overcome the uncertainty, researchers used a probability formula known as Bayes Theorem to assess uncertainty after learning about a crash. Using event recorder data available through a National Highway Traffic Safety Association database, researchers were able to calculate the five characterizing features of several crash events with a degree of certainty. These crash features included the apparent gap the turning vehicle accepted, the speed of the opposing vehicle, the braking rate of the opposing vehicle, and the opposing driver’s reaction time.
In addition to analyzing real-world crashes, researchers tested their calculations with data from instrumented test crashes to confirm that their math was on the mark. “Using the test crash data, we were able to see that our estimates from EDR data seem to be within uncertainty levels,” Davis says. “We plan to do more model checking using instrumented test crashes in the near future.”
Based on their work so far, researchers believe that crash estimation using EDR data is possible, but say some uncertainties remain. “While the precise characterization of individual events continues to be difficult based on EDR data alone, the statistical analyses of large samples is feasible,” Davis says. “As higher resolution data taken at half-second intervals becomes more widely available, that should also allow us to be more precise.”
The research was funded by the U’s Roadway Safety Institute. A seminar of Davis discussing the project is available on the Institute website.
- Research project page
- Roadway Safety Institute seminar: Estimating Drivers’ Behaviors from Event Data Recorder Data, March 22, 2018