Researchers at the Nebraska Transportation Center have developed a new model to help predict when vehicles will change lanes. Their efforts could ultimately help give advanced driver-assistance systems more lead time to react. The researchers detailed the development and results of their model in the journal Transportation Research Record.
If I know the intention, that the vehicle is going to abruptly cut in, I may have a corresponding reaction. I may slow down a little bit, or I may make another lane change to avoid a potential rear-end crash.—Li Zhao, lead author
The team built its model on data from roughly 3,000 vehicles outfitted with front-facing cameras and various sensors. In the early 2010s, the owners of those vehicles drove their regular routes for two months as part of a project funded by the US Department of Transportation—the Safety Pilot Model Deployment (SPMD)—which eventually made the naturalistic driving data available to the public.
Using these data, a lane-changing-based car-following (LCCF) model, which explicitly considers lane-change and car-following behavior simultaneously, is developed. The LCCF model is based on Gaussian-mixture-based hidden Markov model theory and is disaggregated into two processes: LCCF association and LCCF dissociation. These categories are based on the result of the lane change. The overall goal is to predict a driver’s lane-change intention using the LCCF model.—Zhao et al.
To inform the model, Zhao began compiling data from every scenario in which one of those vehicles was following no more than 400 feet, or 3.5 seconds, behind another on the freeway. In some cases, both the front and trailing vehicle were in the same lane before one merged into an adjacent lane; in other cases, one vehicle merged from an adjacent lane, so that both ended up in the same.
Zhao tagged multiple variables that could act as telltales of a driver planning to change lanes: the distance between vehicles, their relative speed, their lateral positions, a subtle turn of the front vehicle’s nose. She then trained a model to analyze the values of those variables at every tenth of a second over a six-second span, from five seconds before a lane change until one second after.
At each of those 60 increments, the model compares the value of each variable—for example, a 10-foot decrease in the distance between vehicles —against the estimated likelihood that the value will occur before a lane change. When all of those variables reach values that indicate the maximum likelihood of a lane change, the model flags the lane change as imminent.
Though it varies a bit across conditions, the model is able to predict a lane change about one second before the center of a vehicle crosses a dividing line on its way to another lane.
One second ahead of time, we start to become confident that the driver is going to make a lane change. That may not mean much to a human driver, but we are talking about automated vehicles or advanced driver-assistance systems. So they can use the lead time to either improve their system, or they can design some extra safety precautions—alerts or warnings, like a crash warning system—to automatically slow down the vehicle or help the driver make some decisions.—Li Zhao
The team picked up on some other interesting trends, too. The average lane change, for example, took between 0.55 and 0.86 seconds. Drivers actually tended to take less time merging into a lane behind a vehicle (the 0.55 seconds) than when changing lanes to get out from behind one (the 0.86). The faster a lane change, the more lead time a driver-assistance system needs in order to compensate, making the distinction a potentially useful one.
Zhao said having access to naturalistic driving data gives her greater confidence in the validity of the model, which she’s optimistic might also be applied to other, richer datasets from vehicles equipped with more sensors and cameras.
Zhao L, Rilett L, Haque MS (2021) “Hidden Markov Model of Lane-Changing-Based Car-Following Behavior on Freeways using Naturalistic Driving Data.” Transportation Research Record doi: 10.1177/0361198121999382