🧬 DriveDNA

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

Recognizing a driver is not the same as capturing driving style — DriveDNA makes vehicle, route, and driving-condition shortcuts measurable.

---

📌 TL;DR

DriveDNA turns a large, in-the-wild naturalistic driving corpus into a benchmark for personalized driving style: representing who is driving as distinct from what they are driving and where. It pairs time-synchronized CAN telemetry (10 Hz) and forward-road video across hundreds of drivers and vehicle models, retains only human-controlled driving (automation-engaged frames removed), and ships a frozen evaluation protocol whose central question is:

Does a model recognize how a person drives — or merely which car they own, which roads they frequent, and which conditions they encounter?

Why it's unique. Public personalized-style resources are small and hold vehicle/route fixed (e.g., PDB: 12 drivers, one car), while large AV datasets (nuScenes, Waymo) carry no persistent driver identity. DriveDNA is the first public corpus combining many drivers × many vehicles × multi-session CAN+video, with clean human-vs-automation separation and explicit confound diagnostics.

✨ Highlights

🧑‍✈️ Drivers	465 persistent de-identified identities (driver_001…), consistent across vehicles
🚗 Vehicle models	115 across 26 brands — 392 drivers share a model with another driver; 22 drivers appear on 2+ models
🛣️ Drives	4,121 decoded drives (Mar 2023 – Jul 2026, multi-continent)
⏱️ Human-controlled driving	975 h total, 581 h in motion, at 10 Hz with forward video
🪟 Benchmark windows	62,674 tagged 60-s windows from 428 drivers (355 in frozen folds)
🏷️ Annotations	6 driving scenarios · 8 behavioral primitives (93.0% audit agreement) · 276,248 maneuver events incl. 22,322 individually verified lane changes
🧪 Protocol	Driver-disjoint splits · 3-seed error bars · frozen evaluation manifests · leakage probes

📡 Modalities & Committed Signals

All streams are decoded from openpilot logs and resampled to a unified 10 Hz grid:

Signal	Meaning	Style construct
vEgo, aEgo (+ jerk)	speed, longitudinal accel	longitudinal aggressiveness
steeringAngleDeg, steeringRateDeg	driver steering INPUT (vehicle-dependent via steer ratio)	steering entropy, reversal rate
actual_curvature	realized path curvature (vehicle-normalized)	cornering sharpness, path geometry
yaw_rate → curv_measured	independently-sensed turning	aggressiveness, slip
leadOne_dRel/vRel/status	lead-vehicle distance & relative speed (radar)	THW, TTC, gap preference
gas, brake (+ pressed)	pedal application (subset of fleet)	pedal dynamics
laneLeft_y, laneRight_y	lane offsets	lane-keeping (SDLP)

Key distinction — steering INPUT vs realized PATH. steeringAngleDeg is the raw wheel input and is vehicle-dependent; actual_curvature is the vehicle-normalized realized path. Their gap is a signal-level handle on the "who vs which-car" question at the heart of the benchmark: vehicle-model probes read 2.3× chance from steering angle but ≈chance from realized curvature.

🎯 Benchmark Tasks & Splits

Task	Input → Output	Metrics
Driver re-identification (core)	k-min support → driver identity	Top-k, AUROC, EER
Personalized behavior prediction (core)	5-s history → 1–5-s future motion	RMSE, PG, MMD/KL/W1
Condition-matched comparison (core)	matched window pair → same driver?	AUROC, EER
Event forecasting (optional)	5-s history → event in 1–5 s	AP, AUROC, lead time
Style explanation (optional)	event window → category + evidence	accuracy (exploratory)

The main driver-disjoint split is 212 train / 45 val / 45 test, plus a 53-driver few-shot hold-out (support and query always from different drives). Additional frozen manifests isolate generalization sources: within-nameplate (same model, different drivers, 24 models), cross-vehicle (same driver, different vehicles), condition-matched pairs (14,868), and missing-channel robustness.

📊 Key Results

Finding	Evidence
Learned representations ≫ classical descriptors	AUROC .935 vs .707 on unseen drivers
Driver signal survives condition matching	.811 ± .006 on 14,868 matched pairs (descriptors → .550, chance)
High re-ID ≠ driving style	Video-only probe hits .937 re-ID but predicts route at 347× chance; collapses to .675 under matching
Recognition ≠ prediction	Best re-ID embedding yields no prediction gain (−0.2%); task-aligned FiLM conditioning does (+0.4 to +1.4%)
Foundation models need adaptation	Zero-shot LLM/TS/VLM rows land at/below the descriptor level; 1-epoch LoRA lifts Qwen3-8B to .871 on event forecasting

30 baseline configurations across five families — representation learning, shortcut robustness, personalization, multimodal modeling, distributional prediction — under one fixed multi-seed protocol.

📦 What's Released

This gated repository contains the full release: de-identified 10 Hz signal tables · raw forward road video (low-resolution 526×330 dashcam view, 1-min segments) · frozen video embeddings (DINOv2/DINOv3/SigLIP2/V-JEPA 2) · all split manifests · VLM scene attributes. The evaluation harness and baseline code live in the GitHub repo.

The signal tables, features, and manifests alone reproduce every number in the paper; video is provided for multimodal research under the same research-only license.

Repository layout (~325 GB):

DriveDNA/
├── data/
│   ├── windows.parquet              # 62,674 benchmark windows: driver, drive, scenario, 8 primitives, stats
│   ├── maneuver_events.parquet      # 276,248 maneuver events (window-indexed)
│   ├── lane_changes_verified.parquet# individually verified lane-change rows
│   ├── vlm_attrs.jsonl / vlm_attrs_qwen3.jsonl   # per-window VLM scene attributes
│   └── t5_llm_subsample.parquet     # frozen 3,000-window event-forecasting subsample
├── features/
│   ├── windows_x.npy                # [62674, 600, 17] 10 Hz CAN windows (+ windows_mask.npy, channels.json)
│   └── windows_vid*.npy             # window-aligned frozen video features (DINOv2/DINOv3/SigLIP2/V-JEPA 2)
├── raw_signals/
│   └── driver_XXX/drive_YYY.parquet # full 10 Hz signal table per drive — 4,121 drives, 460 drivers
├── embeddings/
│   └── {dinov2,dinov3,siglip2,vjepa2}/driver_XXX/drive_YYY.npz   # per-drive frame embeddings
├── videos/
│   └── driver_XXX/drive_YYY/segN.ts # forward road video, 1-min segments (526×330); segN aligns with embeddings
├── splits/                          # driver_folds / cross_route / within_vehicle / cross_vehicle / matched pairs
└── index/
    ├── drives.parquet               # one row per drive: model, human fraction, fold, month, durations
    └── human_segments.parquet       # 12,440 human-driving spans: i0/i1 row range + t0/t1 seconds per drive

🪪 Naming & de-identification

• Drivers are released as driver_001 … driver_460 and each driver's drives as drive_001 … drive_NNN (per-driver, chronological). The raw-ID mapping is held privately by the authors and is append-stable: future data additions never renumber existing entries.
• raw_signals/ tables contain all frames of each drive, including ADAS-engaged frames (cs_enabled / cruiseState_enabled flags included). To reproduce the paper's human-only analysis, slice each drive with index/human_segments:

import pandas as pd

seg = pd.read_parquet("index/human_segments.parquet")          # human-driving spans
s = seg[(seg.driver == "driver_042")].iloc[0]
sig = pd.read_parquet(f"raw_signals/{s.driver}/{s.drive}.parquet")
human = sig.iloc[int(s.i0):int(s.i1)]                          # 10 Hz rows, human-controlled & moving

• index/drives.parquet is the master index: per drive it lists the consolidated vehicle model, human-controlled fraction, minutes of moving human driving, number of human spans, benchmark fold of the driver, and month of recording.

🔒 Ethics & Privacy

• Collected from community drivers with informed consent and compensation; follows source-platform terms.
• Driver and drive identifiers are sequential pseudonyms (driver_001/drive_001); the private raw-ID mapping is withheld by the authors. VINs, device identifiers, precise timestamps, and GPS coordinates removed; no cabin video/audio; forward road video is exterior-only, low-resolution, released under gated research-only access.
• Leakage probes ship as part of the benchmark — users are asked to report leakage alongside utility.
• Prohibited: re-identification attempts; insurance, employment, or law-enforcement scoring of individuals.
• A takedown contact allows any driver to request removal from future versions.

📖 Citation

@article{drivedna2026,
  title   = {DriveDNA: A Large-Scale Multimodal Naturalistic Driving Dataset and
             Benchmark for Driving Style Identification},
  author  = {Wang, Yuhang and Li, Lingyao and Zhou, Hao},
  journal = {arXiv preprint arXiv:2607.23822},
  year    = {2026}
}

🔗 Links & Status

• 📄 Paper: arXiv:2607.23822 · 🤗 Papers page · KDD 2027 Datasets & Benchmarks (under review)
• 💻 Code & harness: github.com/WangYuHang-cmd/DriveDNA
• 📦 Data files: full release (~325 GB) uploading — forward video first, then signal tables, features, and all four embedding families
• ✉️ Contact: haozhou1@usf.edu

---

_{DriveDNA · University of South Florida & University of Arizona · 2026}