🚗 BATON-Sample

Behavioral Analysis of Transition and Operation in Naturalistic Driving

     

A large-scale multimodal benchmark for bidirectional human–DAS control transition in naturalistic driving
Submitted to ACM Multimedia 2026

This is the sample release of BATON — 43 routes covering all 9 modalities, ready for quick exploration and prototyping. For the full 380-route dataset, see HenryYHW/BATON.

---

---

🎬 Live Preview

Continuous sequence — cabin fisheye · front view · 5 fps

8 CAN/IMU sensor streams cycling through all channels

Daytime time-lapse — front · cabin · 2-min intervals

_{Nighttime driving — time-lapse with ⬆ DAS Handover and ↩ Human Takeover event highlights}

---

📦 Sample Release at a Glance

🌍 Routes	👤 Drivers	🚙 Car Models	⏱️ Duration	🔄 Handover Events
43	43	~20	~15 h	~330

🤖 DAS Driving	🧑 Human Driving	⬆️ DAS Handover	↩️ Human Takeover	🌍 Coverage
~52%	~48%	~165	~165	5 Continents

Full dataset: 380 routes · 127 drivers · 84 car models · 136.6 h · 2,892 handover events — available at HenryYHW/BATON

_{Global distribution of participants, per-driver duration, and handover event breakdown (full dataset).}

---

📁 Sample Contents

Each of the 43 routes contains all 9 synchronized modalities:

BATON-Sample/
└── {vehicle_model}/
    └── {driver_id}/
        └── {route_hash}/
            ├── vehicle_dynamics.csv   # Speed, accel, steering, pedals, DAS status
            ├── planning.csv           # DAS curvature, lane change intent
            ├── radar.csv              # Lead vehicle distance & relative speed
            ├── driver_state.csv       # Face pose, eye openness, awareness
            ├── imu.csv                # 3-axis accel & gyro at 100 Hz
            ├── gps.csv                # Coordinates, heading
            ├── localization.csv       # Road curvature, lane position
            ├── qcamera.mp4            # Front-view video (526×330, H.264, 20 fps)
            └── dcamera.mp4            # In-cabin fisheye video (1928×1208, HEVC, 20 fps)

---

🔬 Data Collection & Modalities

Setup: Non-intrusive plug-and-play OBD-II dongle + dual cameras. Drivers use their own vehicles during real daily commutes — no lab, no script. Component Spec 📡 OBD-II Dongle CAN-bus at 100 Hz 📷 Front camera 526×330 · H.264 · 20 fps 🎥 Cabin fisheye 1928×1208 · HEVC · 20 fps 🛰️ GPS 10 Hz 9 synchronized modalities: vehicle_dynamics.csv — speed, accel, steering, pedals, DAS status planning.csv — DAS curvature, lane change intent radar.csv — lead vehicle distance & relative speed driver_state.csv — face pose, eye openness, awareness imu.csv — 3-axis accel & gyro at 100 Hz gps.csv — coordinates, heading localization.csv — road curvature, lane position qcamera.mp4 — front-view video dcamera.mp4 — in-cabin fisheye video

📷 Front · Day 📷 Front · Night ⬆️ DAS Handover 🎥 Cabin · Day 🎥 Cabin · Night ↩️ Human Takeover

_{Aligned multimodal streams around a HANDOVER event: cabin video · front video · GPS trajectory · sensor signals.}

---

🏆 Benchmark Tasks

Task	Description	Samples (full)	Labels	Primary Metric
🎯 Task 1	Driving action recognition (7-class)	979,809	Cruising · Car Following · Accelerating · Braking · Lane Change · Turning · Stopped	Macro-F1
⬆️ Task 2	Handover prediction (Human→DAS)	56,564	Handover (14.9%) · No Handover	AUPRC
↩️ Task 3	Takeover prediction (DAS→Human)	71,079	Takeover (11.9%) · No Takeover	AUPRC

Evaluation protocol: Cross-driver split · 5-second input window · 3-second prediction horizon · 3 seeds (42, 123, 7)

---

🚀 Quick Start

1. Get the sample data

# Clone this sample dataset (~few GB, all modalities, 43 routes)
git lfs install
git clone https://huggingface.co/datasets/HenryYHW/BATON-Sample

# Or via Python
from huggingface_hub import snapshot_download
snapshot_download('HenryYHW/BATON-Sample', repo_type='dataset', local_dir='./data')

2. Get the full dataset

# Full dataset (380 routes) — requires HuggingFace account
python -c "
from huggingface_hub import snapshot_download
snapshot_download('HenryYHW/BATON', repo_type='dataset', local_dir='./data')
"

3. Extract video features

cd data_processing

# EfficientNet-B0 features (used in main baselines)
python extract_front_video_features.py
python extract_cabin_video_features.py

4. Train baselines (requires full dataset + benchmark files)

cd baseline

# GRU on all modalities — Task 1
python train_nn.py --task task1 --modality Full-All --model gru --seed 42

# XGBoost on structured signals — Task 2
python train_classical.py --task task2 --model xgb --seed 42

# Zero-shot VLM baseline (GPT-4o or Gemini 2.5 Flash)
python run_vlm.py --model gpt4o --task task1

See GitHub — OpenLKA/BATON for the complete codebase.

---

📐 Evaluation Protocol

Setting	Value
Primary split	Cross-driver (disjoint drivers in train / val / test)
Additional splits	Cross-vehicle, Random
Input window	5 seconds
Prediction horizon	1 s, 3 s, 5 s (main: 3 s)
Random seeds	42, 123, 7 — report 3-seed average
Task 1 metric	Macro-F1
Task 2 / 3 metrics	AUPRC (primary), AUC-ROC, F1

---

📡 Data Access

Resource	Link
🔍 This Sample (43 routes)	HuggingFace — HenryYHW/BATON-Sample
📦 Full Dataset (380 routes)	HuggingFace — HenryYHW/BATON
💻 Code & Baselines	GitHub — OpenLKA/BATON
📄 arXiv Paper	arxiv.org/abs/2604.07263

---

📜 Citation

@article{wang2026baton,
  title   = {BATON: A Multimodal Benchmark for Bidirectional Automation Transition
             Observation in Naturalistic Driving},
  author  = {Wang, Yuhang and Xu, Yiyao and Yang, Chaoyun and Li, Lingyao
             and Sun, Jingran and Zhou, Hao},
  journal = {arXiv preprint arXiv:2604.07263},
  year    = {2026}
}

---

📄 License

This dataset is released for academic research use only under CC BY-NC 4.0 (Creative Commons Attribution–NonCommercial 4.0 International).

You are free to use and redistribute the data for non-commercial research, and to adapt or build upon it for non-commercial purposes — provided that:

• Attribution — You must cite the BATON paper (see Citation above) in any publication or work that uses this dataset.
• Non-Commercial — Commercial use of this dataset or any derivative is strictly prohibited.
• Academic Use Only — This dataset is intended solely for academic research. Use in any commercial product, service, or application is not permitted.

For commercial licensing inquiries, please contact the authors.

---

_{🔗 Paper  ·  Full Dataset  ·  Sample Dataset  ·  GitHub}