DriveDNA: A Large-Scale Multimodal Naturalistic Driving Dataset and Benchmark for Driving Style Identification

Researchers introduced DriveDNA, a naturalistic multimodal driving dataset with 4,121 drives from 465 drivers across 115 vehicle models.

The benchmark covers few-shot driver re-identification, personalized behavior prediction, and condition-matched comparison, with behavioral annotations and 276,248 maneuver events.

Results show learned representations beat classical descriptors, but video-only models can appear stronger by exploiting route leakage.

DriveDNA helps test whether models capture true driver behavior instead of shortcuts from the vehicle, route, or scene.