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

Researchers released DriveDNA, a large naturalistic driving dataset with 4,121 real-world drives from 465 drivers across 115 vehicle models.

The benchmark tests personalized driving-style modeling through driver re-identification, behavior prediction, and condition-matched comparison tasks.

It uses multimodal data, including forward video, to better capture real driving behavior over 975 hours of driving.

Early results suggest models can look strong by exploiting route or vehicle context, so robust evaluation is critical.