September 9, 2025 by Patricia DeLacey, University of Michigan
Collected at: https://techxplore.com/news/2025-09-hidden-insights-gps-track-lane.html
Understanding how and when drivers change lanes is key to improving highway traffic flow, safety and autonomous vehicle performance, and a new approach developed at the University of Michigan outperforms current methods using only GPS data.
Up to this point, lane change estimation has been done using on-board cameras or lane-level high-resolution maps that provide geometry, lane markings and lane connections. Both methods are expensive and not always reliable. Cameras fail when the lane lines are faded or occluded and maps are difficult to update at a large scale.
“Almost every car sold in the last decade collects GPS data, but until now, you could not use that data without maps to understand driver behavior, like aggressiveness or how people respond to traffic conditions,” said Arpan Kusari, an assistant research scientist at the University of Michigan Transportation Research Institute (UMTRI) and senior author of a study on the work published in IEEE Transactions on Intelligent Transportation Systems.
“By leveraging this stream of information, we’ve developed a method that can reliably detect lane changes, offering a simple, cost-effective way to study driving behavior and improve both traffic safety and autonomous vehicle performance.”
Unlocking ‘dark’ GPS trajectory data
The researchers used real vehicle trajectories from 130 participants in UMTRI’s Safety Pilot Model Deployment project. To tap into the wealth of available GPS trajectory data, they introduced a new, nonparametric unsupervised approach that completely departs from current methods, Kusari said. The approach conceptualizes each vehicle’s GPS trajectory as two elements: large-scale variations which correspond to road curvature and small-scale variations that predominantly represent lane changes.
To tease apart the small-scale variations from the large-scale ones, each trajectory is transformed using mathematical functions that analyze different stretches of the driving path—between 100 and 500 meters—to calculate how much each point deviates from the expected path. The transformed data, called the Basis-Aligned Coordinate System, or BACS, helps lane changes pop out as obvious departures from normal highway driving patterns.
“This innovation not only unlocks the potential of massive amounts of previously unusable ‘dark’ trajectory data for applications like advanced driver assistance or automated driving system labeling, but also provides a computationally simple and highly robust solution where previous methods fall short,” said Kusari.
The BACS method is immensely generalizable and can be applied to virtually all manner of GPS trajectories without any restrictions to geography.
To validate the new method, the researchers compared lane changes picked out by BACS from GPS trajectories against map-based and camera-based ground truths. BACS proved highly effective at capturing lane change events, with fewer false negatives than camera-based benchmarks and performing on par with map-based ones.
While the BACS method does not use the yaw—rotation around the vehicle’s vertical axis that is particularly apparent on curves—it can approximate the yaw very closely. This serves as another validation that the mathematical model is appropriately modeling physical space.
BACS did have a higher number of false positives, particularly at clover leaf joints, stemming from unexpected road curvatures like sharp kinks that the system mistook for lane changes. To reduce false positives in future versions, the researchers plan to analyze multiple vehicle trajectories on the same road segments, which should help distinguish between genuine lane changes and road geometry that tricks the system.
How GPS data could improve safety
Overall, this cost-effective method could help label lane changes in naturalistic highway driving datasets, like those collected by UMTRI. From here, follow-up analyses could provide a deeper understanding of driver lane-change patterns while identifying risky behaviors. Policymakers could then leverage that information to improve traffic and road safety. AV developers could use it to improve how vehicles predict nearby cars’ lane changes, and to gain deeper insight into the behavior of the human they’re sharing the wheel with.
For example, this method could offer an inexpensive alternative for driver state monitoring in AVs, which typically relies on an in-cabin camera to assess whether the driver’s eyes are on the road. Vehicles could potentially provide real-time alerts based on the frequency of lane changes and swerves, Kusari said.
It could also be used to analyze how the time of day—a proxy for fatigue—and lighting conditions affects driving styles at a group-level (protecting individual privacy.) Insights from anonymized driving data could inform regulations, safety campaigns and algorithms that anticipate risky human driving behaviors and adjust the vehicle’s responses.
“Our newly developed nonparametric unsupervised approach for robust lane change estimation fundamentally shifts how we analyze driving behavior and utilize data in intelligent transportation systems,” said Kusari.
More information: Manav Prabhakar et al, A Non-Parametric Unsupervised Approach for Robust Lane-Change Estimation From Highway-Based Trajectory Data, IEEE Transactions on Intelligent Transportation Systems (2025). DOI: 10.1109/TITS.2025.3591063
Journal information: IEEE Transactions on Intelligent Transportation Systems
Leave a Reply