HenryYHW commited on
Commit
e7d8a1f
·
verified ·
1 Parent(s): ab2b5de

Update dataset card: rename to ADAS-TO, update GIF labels, anonymous authors

Browse files
Files changed (1) hide show
  1. README.md +423 -424
README.md CHANGED
@@ -1,424 +1,423 @@
1
- ---
2
- license: cc-by-nc-4.0
3
- task_categories:
4
- - time-series-forecasting
5
- - video-classification
6
- tags:
7
- - autonomous-driving
8
- - ADAS
9
- - takeover
10
- - driver-behavior
11
- - time-series
12
- - multimodal
13
- - CAN-bus
14
- - vehicle-dynamics
15
- - driving-safety
16
- - human-factors
17
- size_categories:
18
- - 10K<n<100K
19
- language:
20
- - en
21
- pretty_name: >-
22
- ADAS-TO: A Large-Scale Multimodal Naturalistic Dataset and Empirical
23
- Characterization of Human Takeovers during ADAS Engagement
24
- ---
25
-
26
- <div align="center">
27
-
28
- # 🚗💨 ADAS-TO
29
-
30
- ### **The Largest Naturalistic ADAS Takeover Dataset**
31
-
32
- *15,705 real-world takeover events · 327 drivers · 163 vehicle models · 23 manufacturers*
33
-
34
- [![License: CC BY-NC 4.0](https://img.shields.io/badge/License-CC%20BY--NC%204.0-lightgrey.svg)](https://creativecommons.org/licenses/by-nc/4.0/)
35
- [![Dataset on HF](https://img.shields.io/badge/🤗%20Dataset-HADAS--TakeOver-blue)](https://huggingface.co/datasets/HenryYHW/ADAS-TO)
36
- [![Clips](https://img.shields.io/badge/Clips-15%2C705-brightgreen)]()
37
- [![Vehicles](https://img.shields.io/badge/Vehicle%20Models-163-orange)]()
38
- [![Size](https://img.shields.io/badge/Size-~33%20GB-red)]()
39
- [![Modality](https://img.shields.io/badge/Modality-Video%20%2B%20CAN%20%2B%20Radar%20%2B%20IMU-purple)]()
40
-
41
- ---
42
-
43
- **When does a human driver take over from an ADAS?** **Why?** **How?**
44
-
45
- ADAS-TO captures the critical moment of control transition — the exact instant a driver decides the automation is no longer sufficient — across thousands of real-world scenarios with synchronized front-view video, vehicle dynamics, radar, and IMU data.
46
-
47
- </div>
48
-
49
- ---
50
-
51
- ## 🎬 Takeover Examples
52
-
53
- <div align="center">
54
-
55
- *Each GIF shows ±3 seconds around the takeover moment — ADAS engaged → driver takes control*
56
-
57
- <table>
58
- <tr>
59
- <td align="center"><img src="https://huggingface.co/datasets/HenryYHW/ADAS-TO/resolve/main/assets/takeover_1.gif" width="240"/><br/><sub>Highway Scenario</sub></td>
60
- <td align="center"><img src="https://huggingface.co/datasets/HenryYHW/ADAS-TO/resolve/main/assets/takeover_2.gif" width="240"/><br/><sub>Urban Driving</sub></td>
61
- <td align="center"><img src="https://huggingface.co/datasets/HenryYHW/ADAS-TO/resolve/main/assets/takeover_3.gif" width="240"/><br/><sub>Traffic Interaction</sub></td>
62
- <td align="center"><img src="https://huggingface.co/datasets/HenryYHW/ADAS-TO/resolve/main/assets/takeover_4.gif" width="240"/><br/><sub>Complex Road</sub></td>
63
- </tr>
64
- <tr>
65
- <td align="center"><img src="https://huggingface.co/datasets/HenryYHW/ADAS-TO/resolve/main/assets/takeover_5.gif" width="240"/><br/><sub>Lane Change</sub></td>
66
- <td align="center"><img src="https://huggingface.co/datasets/HenryYHW/ADAS-TO/resolve/main/assets/takeover_6.gif" width="240"/><br/><sub>Multi-Vehicle</sub></td>
67
- <td align="center"><img src="https://huggingface.co/datasets/HenryYHW/ADAS-TO/resolve/main/assets/takeover_7.gif" width="240"/><br/><sub>Suburban Road</sub></td>
68
- <td align="center"><img src="https://huggingface.co/datasets/HenryYHW/ADAS-TO/resolve/main/assets/takeover_8.gif" width="240"/><br/><sub>Dynamic Scene</sub></td>
69
- </tr>
70
- </table>
71
-
72
- </div>
73
-
74
- ---
75
-
76
- ## 📊 Dataset at a Glance
77
-
78
- <div align="center">
79
-
80
- | | Statistic | Value |
81
- |:---:|:---|:---|
82
- | 🎥 | **Total takeover clips** | **15,705** |
83
- | 👤 | **Unique drivers** | **327** |
84
- | 🛣️ | **Unique driving routes** | **2,312** |
85
- | 🚘 | **Vehicle models** | **163** |
86
- | 🏭 | **Manufacturers** | **23** |
87
- | ⏱️ | **Clip duration** | **20 seconds** (±10s around takeover) |
88
- | 📹 | **Video** | Front-facing camera, **20 fps** |
89
- | 📡 | **CAN / sensor signals** | **10–100 Hz** |
90
- | 📁 | **Files per clip** | **10** (1 video + 1 meta + 8 CSV) |
91
- | 💾 | **Total size** | **~33 GB** |
92
-
93
- </div>
94
-
95
- ---
96
-
97
- ## 🔥 Why ADAS-TO?
98
-
99
- > *"The takeover moment is the most safety-critical instant in human-automation interaction — yet it remains one of the least studied due to lack of data."*
100
-
101
- ### 🏆 Unprecedented Scale
102
- Over **15,000 real-world takeover events** — orders of magnitude larger than existing datasets that typically contain hundreds of events captured in driving simulators.
103
-
104
- ### 🌍 Unmatched Diversity
105
- **163 vehicle models** from **23 manufacturers** including Tesla, Toyota, Honda, Hyundai, Ford, Volkswagen, Rivian, and more. From compact EVs to full-size trucks — spanning the full spectrum of modern ADAS implementations.
106
-
107
- ### 🎯 Rich Multimodal Signals
108
- Every clip contains **synchronized** front-camera video, vehicle dynamics (speed, acceleration, steering), ADAS controller state, control commands, actuator outputs, driving model predictions, radar/lead vehicle data, and IMU measurements.
109
-
110
- ### 🌐 Real-World Naturalistic Data
111
- Collected through **online and offline autonomous driving communities** with diverse real-world driving conditions — highways, urban streets, suburbs, varying weather and lighting. No simulators. No scripted scenarios. Pure naturalistic driving behavior.
112
-
113
- ---
114
-
115
- ## 🎯 Use Cases
116
-
117
- | Application | Description |
118
- |:---|:---|
119
- | 🔮 **Takeover Prediction** | Build early warning systems that predict when a driver will need to take over |
120
- | 🧠 **Driver Behavior Modeling** | Understand human responses during control transitions |
121
- | 📈 **ADAS Performance Analysis** | Compare disengagement patterns across vehicle types and ADAS systems |
122
- | 🤖 **Autonomous Driving Safety** | Train and evaluate safety-critical decision-making models |
123
- | 🧪 **Human Factors Research** | Study cognitive load, reaction times, and situational awareness |
124
- | 📊 **Multimodal Time-Series** | Develop forecasting and classification models on rich temporal data |
125
- | 🏗️ **HMI Design** | Design better human-machine interfaces for automated vehicles |
126
-
127
- ---
128
-
129
- ## 📁 Dataset Structure
130
-
131
- ```
132
- ADAS-TO/
133
- ── <CAR_MODEL>/ # e.g., TOYOTA_PRIUS, TESLA_AP3_MODEL_3
134
- └── <driver_XXX>/ # 🔒 anonymized driver ID
135
- ── <route_XXX>/ # 🔒 anonymized route ID
136
- ── <clip_id>/ # integer (0-indexed per route)
137
- │ ├── 🎥 takeover.mp4 20-second front-camera video
138
- │ ├── 📋 meta.json clip metadata & timing
139
- │ ├── 🚗 carState.csv vehicle dynamics & driver inputs
140
- │ ├── 🤖 controlsState.csv ADAS controller state & alerts
141
- │ ├── 🎮 carControl.csv lateral/longitudinal commands
142
- │ ├── ⚙️ carOutput.csv actuator outputs
143
- │ ├── 🧠 drivingModelData.csv model predictions & lane detection
144
- ── 📡 radarState.csv lead vehicle radar data
145
- │ ├── 📐 accelerometer.csv IMU acceleration data
146
- │ └── 📏 longitudinalPlan.csv planner targets & FCW
147
- └── ...
148
- ```
149
-
150
- ---
151
-
152
- ## 📐 Takeover Event Definition
153
-
154
- <div align="center">
155
-
156
- ```
157
- ◄──────── 10 seconds ────────►◄──────── 10 seconds ────────►
158
- ┌──────────────────────────────┬──────────────────────────────┐
159
- │ 🤖 ADAS ENGAGED │ 👤 MANUAL CONTROL │
160
- │ (automation driving) │ (driver takes over) │
161
- └──────────────────────────────┴──────────────────────────────┘
162
-
163
- TAKEOVER EVENT
164
- (ON → OFF transition)
165
- ```
166
-
167
- </div>
168
-
169
- A **takeover event** is detected as an ADAS ON → OFF transition satisfying:
170
-
171
- | Criterion | Value |
172
- |:---|:---|
173
- | **ADAS engaged** | `controlsState.enabled` OR `cruiseState.enabled` |
174
- | **Min ON duration** | 2 seconds before disengagement |
175
- | **Min OFF duration** | 2 seconds after disengagement |
176
- | **Gap merging** | Transient gaps < 0.5s merged (filters sensor noise) |
177
- | **Clip window** | ±10 seconds centered on transition (20s total) |
178
-
179
- ---
180
-
181
- ## 📑 Data Fields Reference
182
-
183
- ### 📋 meta.json Clip Metadata
184
-
185
- | Field | Type | Description |
186
- |:---|:---|:---|
187
- | `car_model` | string | Vehicle model (e.g., `TOYOTA_PRIUS`) |
188
- | `dongle_id` | string | Anonymized driver ID (`driver_XXX`) |
189
- | `route_id` | string | Anonymized route ID (`route_XXX`) |
190
- | `log_kind` | string | Log resolution: `qlog` (10 Hz) or `rlog` (100 Hz) |
191
- | `log_hz` | int | CAN signal sampling rate |
192
- | `vid_kind` | string | Camera source type |
193
- | `camera_fps` | int | Video frame rate (20 fps) |
194
- | `clip_id` | int | Clip index within route (0-indexed) |
195
- | `event_mono` | int | Monotonic timestamp of takeover (ns) |
196
- | `video_time_s` | float | Takeover time within full route video (s) |
197
- | `clip_start_s` | float | Clip start time within route (s) |
198
- | `clip_dur_s` | float | Clip duration (s) |
199
-
200
- ### 🚗 carState.csv Vehicle Dynamics & Driver Inputs
201
-
202
- | Column | Unit | Description |
203
- |:---|:---|:---|
204
- | `vEgo` | m/s | Ego vehicle speed |
205
- | `aEgo` | m/s² | Ego vehicle acceleration |
206
- | `steeringAngleDeg` | deg | Steering wheel angle |
207
- | `steeringTorque` | N·m | Driver steering torque |
208
- | `steeringPressed` | bool | Driver actively steering |
209
- | `gasPressed` | bool | Gas pedal pressed |
210
- | `brakePressed` | bool | Brake pedal pressed |
211
- | `cruiseState.enabled` | bool | Cruise / ADAS engaged |
212
-
213
- ### 🤖 controlsState.csv ADAS Controller
214
-
215
- | Column | Unit | Description |
216
- |:---|:---|:---|
217
- | `enabled` | bool | ADAS system enabled |
218
- | `active` | bool | ADAS actively controlling vehicle |
219
- | `curvature` | 1/m | Current path curvature |
220
- | `desiredCurvature` | 1/m | Target curvature from planner |
221
- | `vCruise` | m/s | Set cruise speed |
222
- | `longControlState` | enum | Longitudinal control state |
223
- | `alertText1` | string | Primary driver alert |
224
- | `alertText2` | string | Secondary driver alert |
225
-
226
- ### 🎮 carControl.csv Control Commands
227
-
228
- | Column | Unit | Description |
229
- |:---|:---|:---|
230
- | `latActive` | bool | Lateral control active |
231
- | `longActive` | bool | Longitudinal control active |
232
- | `actuators.accel` | m/ | Commanded acceleration |
233
- | `actuators.torque` | N·m | Commanded steering torque |
234
- | `actuators.curvature` | 1/m | Commanded path curvature |
235
-
236
- ### ⚙️ carOutput.csv Actuator Outputs
237
-
238
- | Column | Description |
239
- |:---|:---|
240
- | `actuatorsOutput.accel` | Acceleration actuator output |
241
- | `actuatorsOutput.brake` | Brake actuator output |
242
- | `actuatorsOutput.gas` | Gas actuator output |
243
- | `actuatorsOutput.steer` | Steering actuator output |
244
- | `actuatorsOutput.steerOutputCan` | Raw CAN steering output |
245
- | `actuatorsOutput.steeringAngleDeg` | Steering angle output (deg) |
246
-
247
- ### 🧠 drivingModelData.csv Driving Model Predictions
248
-
249
- | Column | Description |
250
- |:---|:---|
251
- | `action.desiredCurvature` | Model-predicted desired curvature |
252
- | `action.desiredAcceleration` | Model-predicted desired acceleration |
253
- | `laneLineMeta.leftProb` | Left lane line detection probability |
254
- | `laneLineMeta.rightProb` | Right lane line detection probability |
255
-
256
- ### 📡 radarState.csv Lead Vehicle Detection
257
-
258
- | Column | Unit | Description |
259
- |:---|:---|:---|
260
- | `leadOne.dRel` | m | Distance to primary lead vehicle |
261
- | `leadOne.vRel` | m/s | Relative velocity of lead |
262
- | `leadOne.vLead` | m/s | Absolute velocity of lead |
263
- | `leadOne.aLeadK` | m/s² | Lead vehicle acceleration |
264
- | `leadTwo.*` | | Secondary lead vehicle (same fields) |
265
-
266
- ### 📐 accelerometer.csv IMU Data
267
-
268
- | Column | Unit | Description |
269
- |:---|:---|:---|
270
- | `acceleration.v` | m/s² | 3-axis acceleration vector |
271
- | `timestamp` || Sensor timestamp |
272
-
273
- ### 📏 longitudinalPlan.csv Planner Outputs
274
-
275
- | Column | Unit | Description |
276
- |:---|:---|:---|
277
- | `aTarget` | m/s² | Target acceleration |
278
- | `hasLead` | bool | Lead vehicle detected |
279
- | `fcw` | bool | Forward collision warning active |
280
- | `speeds[]` | m/s | Planned speed profile |
281
- | `accels[]` | m/s² | Planned acceleration profile |
282
-
283
- ---
284
-
285
- ## 🚘 Vehicle Coverage
286
-
287
- <div align="center">
288
-
289
- **23 Manufacturers · 163 Models · From Compact EVs to Full-Size Trucks**
290
-
291
- </div>
292
-
293
- ### Top Vehicle Models by Clip Count
294
-
295
- | # | Vehicle Model | Clips | | # | Vehicle Model | Clips |
296
- |:---:|:---|---:|:---:|:---:|:---|---:|
297
- | 1 | 🏆 RIVIAN R1 GEN1 | 2,127 | | 10 | CHEVROLET BOLT EUV | 244 |
298
- | 2 | 🥈 ACURA MDX 3G | 1,863 | | 11 | TOYOTA RAV4 TSS2 | 228 |
299
- | 3 | 🥉 FORD F-150 MK14 | 1,226 | | 12 | RAM HD 5TH GEN | 221 |
300
- | 4 | CHEVROLET SILVERADO | 639 | | 13 | VOLKSWAGEN JETTA MK7 | 215 |
301
- | 5 | TOYOTA PRIUS | 482 | | 14 | KIA EV6 | 209 |
302
- | 6 | HONDA CIVIC | 470 | | 15 | VOLKSWAGEN GOLF MK7 | 192 |
303
- | 7 | TESLA MODEL 3 | 432 | | 16 | KIA NIRO EV | 185 |
304
- | 8 | FORD MAVERICK MK1 | 300 | | 17 | HYUNDAI IONIQ 6 | 177 |
305
- | 9 | HYUNDAI IONIQ 5 | 266 | | 18 | VOLKSWAGEN ATLAS MK1 | 153 |
306
-
307
- <details>
308
- <summary>📋 <b>All 23 Manufacturers</b> (click to expand)</summary>
309
-
310
- > Acura · Audi · BYD · Chevrolet · Ford · Genesis · Honda · Hyundai · Jeep · Kia · Lexus · Mazda · Nissan · Porsche · RAM · Rivian · Skoda · Subaru · Tesla · Toyota · Volkswagen · Volvo
311
-
312
- </details>
313
-
314
- ---
315
-
316
- ## 🚀 Quick Start
317
-
318
- ### Loading a Single Clip
319
-
320
- ```python
321
- import json
322
- import pandas as pd
323
- from huggingface_hub import hf_hub_download
324
-
325
- repo_id = "HenryYHW/ADAS-TO"
326
- clip_path = "TOYOTA_PRIUS/driver_001/route_001/0"
327
-
328
- # 📋 Download metadata
329
- meta_path = hf_hub_download(repo_id, f"{clip_path}/meta.json", repo_type="dataset")
330
- with open(meta_path) as f:
331
- meta = json.load(f)
332
-
333
- # 🚗 Load vehicle state signals
334
- car_state = pd.read_csv(
335
- hf_hub_download(repo_id, f"{clip_path}/carState.csv", repo_type="dataset")
336
- )
337
- print(car_state[["vEgo", "aEgo", "steeringAngleDeg", "brakePressed"]].describe())
338
-
339
- # 🤖 Load ADAS controller state
340
- controls = pd.read_csv(
341
- hf_hub_download(repo_id, f"{clip_path}/controlsState.csv", repo_type="dataset")
342
- )
343
-
344
- # 📡 Load radar data
345
- radar = pd.read_csv(
346
- hf_hub_download(repo_id, f"{clip_path}/radarState.csv", repo_type="dataset")
347
- )
348
- ```
349
-
350
- ### Iterating Over All Clips
351
-
352
- ```python
353
- from huggingface_hub import HfApi
354
-
355
- api = HfApi()
356
- files = api.list_repo_files("HenryYHW/ADAS-TO", repo_type="dataset")
357
- meta_files = [f for f in files if f.endswith("meta.json")]
358
- print(f"Total clips: {len(meta_files)}") # → 15,705
359
- ```
360
-
361
- ### 💾 Download the Full Dataset
362
-
363
- ```bash
364
- # Using huggingface-cli (recommended)
365
- huggingface-cli download HenryYHW/ADAS-TO --repo-type dataset --local-dir ./ADAS-TO
366
-
367
- # Using git-lfs
368
- git lfs install
369
- git clone https://huggingface.co/datasets/HenryYHW/ADAS-TO
370
- ```
371
-
372
- ---
373
-
374
- ## 🔒 Privacy & Ethics
375
-
376
- - **Anonymized identifiers**: All driver and route IDs are replaced with anonymous tokens (`driver_XXX`, `route_XXX`)
377
- - **Forward-view only**: Video captures road-facing view only no cabin or driver footage
378
- - **No PII**: No personally identifiable information is included in any data file
379
- - **Community-sourced**: Data collected through autonomous driving enthusiast communities with informed participation
380
-
381
- ---
382
-
383
- ## 📖 Data Collection
384
-
385
- ADAS-TO was built from naturalistic driving logs contributed by **online and offline autonomous driving communities**. Participating drivers voluntarily shared their driving data collected through various ADAS-equipped vehicles during everyday driving. The raw logs were processed through an automated pipeline to:
386
-
387
- 1. **Detect** ADAS disengagement events (ON→OFF transitions)
388
- 2. **Extract** synchronized video and CAN-bus signals within a 20-second window
389
- 3. **Validate** each clip for signal completeness and temporal alignment
390
- 4. **Anonymize** all driver and route identifiers
391
-
392
- This community-driven collection approach enables unprecedented scale and diversity, capturing genuine driver behavior across a wide spectrum of vehicles, road types, and driving conditions.
393
-
394
- ---
395
-
396
- ## 📝 Citation
397
-
398
- If you use ADAS-TO in your research, please cite:
399
-
400
- ```bibtex
401
- @dataset{adas-to_2026,
402
- title = {ADAS-TO: A Large-Scale Multimodal Naturalistic Dataset and Empirical Characterization of Human Takeovers during ADAS Engagement},
403
- author = {Anonymous Authors},
404
- year = {2026},
405
- publisher = {Hugging Face},
406
- url = {https://huggingface.co/datasets/HenryYHW/ADAS-TO}
407
- }
408
- ```
409
-
410
- ---
411
-
412
- ## 📄 License
413
-
414
- <div align="center">
415
-
416
- This dataset is released under [**CC BY-NC 4.0**](https://creativecommons.org/licenses/by-nc/4.0/).
417
-
418
- For academic and non-commercial research purposes.
419
-
420
- ---
421
-
422
- *Built with ❤️ for the autonomous driving research community*
423
-
424
- </div>
 
1
+ ---
2
+ license: cc-by-nc-4.0
3
+ task_categories:
4
+ - time-series-forecasting
5
+ - video-classification
6
+ tags:
7
+ - autonomous-driving
8
+ - ADAS
9
+ - takeover
10
+ - driver-behavior
11
+ - time-series
12
+ - multimodal
13
+ - CAN-bus
14
+ - vehicle-dynamics
15
+ - driving-safety
16
+ - human-factors
17
+ size_categories:
18
+ - 10K<n<100K
19
+ language:
20
+ - en
21
+ pretty_name: "ADAS-TO: A Large-Scale Multimodal Naturalistic Dataset and Empirical Characterization of Human Takeovers during ADAS Engagement"
22
+ ---
23
+
24
+ <div align="center">
25
+
26
+ # 🚗💨 ADAS-TO
27
+
28
+ ### **A Large-Scale Multimodal Naturalistic Dataset and Empirical Characterization of Human Takeovers during ADAS Engagement**
29
+
30
+ *15,705 real-world takeover events · 327 drivers · 163 vehicle models · 23 manufacturers*
31
+
32
+ [![License: CC BY-NC 4.0](https://img.shields.io/badge/License-CC%20BY--NC%204.0-lightgrey.svg)](https://creativecommons.org/licenses/by-nc/4.0/)
33
+ [![Dataset on HF](https://img.shields.io/badge/🤗%20Dataset-ADAS--TO-blue)](https://huggingface.co/datasets/HenryYHW/ADAS-TO)
34
+ [![Clips](https://img.shields.io/badge/Clips-15%2C705-brightgreen)]()
35
+ [![Vehicles](https://img.shields.io/badge/Vehicle%20Models-163-orange)]()
36
+ [![Size](https://img.shields.io/badge/Size-~33%20GB-red)]()
37
+ [![Modality](https://img.shields.io/badge/Modality-Video%20%2B%20CAN%20%2B%20Radar%20%2B%20IMU-purple)]()
38
+
39
+ ---
40
+
41
+ **When does a human driver take over from an ADAS?** **Why?** **How?**
42
+
43
+ ADAS-TO captures the critical moment of control transition — the exact instant a driver decides the automation is no longer sufficient — across thousands of real-world scenarios with synchronized front-view video, vehicle dynamics, radar, and IMU data.
44
+
45
+ </div>
46
+
47
+ ---
48
+
49
+ ## 🎬 Takeover Examples
50
+
51
+ <div align="center">
52
+
53
+ *Each GIF shows ±3 seconds around the takeover moment — ADAS engaged → driver takes control*
54
+
55
+ <table>
56
+ <tr>
57
+ <td align="center"><img src="https://huggingface.co/datasets/HenryYHW/ADAS-TO/resolve/main/assets/takeover_1.gif" width="240"/><br/><sub>On-coming Traffic</sub></td>
58
+ <td align="center"><img src="https://huggingface.co/datasets/HenryYHW/ADAS-TO/resolve/main/assets/takeover_2.gif" width="240"/><br/><sub>Bridge</sub></td>
59
+ <td align="center"><img src="https://huggingface.co/datasets/HenryYHW/ADAS-TO/resolve/main/assets/takeover_3.gif" width="240"/><br/><sub>Night Driving</sub></td>
60
+ <td align="center"><img src="https://huggingface.co/datasets/HenryYHW/ADAS-TO/resolve/main/assets/takeover_4.gif" width="240"/><br/><sub>Sharp Curve</sub></td>
61
+ </tr>
62
+ <tr>
63
+ <td align="center"><img src="https://huggingface.co/datasets/HenryYHW/ADAS-TO/resolve/main/assets/takeover_5.gif" width="240"/><br/><sub>Surrounding Car</sub></td>
64
+ <td align="center"><img src="https://huggingface.co/datasets/HenryYHW/ADAS-TO/resolve/main/assets/takeover_6.gif" width="240"/><br/><sub>Traffic Light</sub></td>
65
+ <td align="center"><img src="https://huggingface.co/datasets/HenryYHW/ADAS-TO/resolve/main/assets/takeover_7.gif" width="240"/><br/><sub>Lane Change</sub></td>
66
+ <td align="center"><img src="https://huggingface.co/datasets/HenryYHW/ADAS-TO/resolve/main/assets/takeover_8.gif" width="240"/><br/><sub>Hard Brake</sub></td>
67
+ </tr>
68
+ </table>
69
+
70
+ </div>
71
+
72
+ ---
73
+
74
+ ## 📊 Dataset at a Glance
75
+
76
+ <div align="center">
77
+
78
+ | | Statistic | Value |
79
+ |:---:|:---|:---|
80
+ | 🎥 | **Total takeover clips** | **15,705** |
81
+ | 👤 | **Unique drivers** | **327** |
82
+ | 🛣️ | **Unique driving routes** | **2,312** |
83
+ | 🚘 | **Vehicle models** | **163** |
84
+ | 🏭 | **Manufacturers** | **23** |
85
+ | ⏱️ | **Clip duration** | **20 seconds** (±10s around takeover) |
86
+ | 📹 | **Video** | Front-facing camera, **20 fps** |
87
+ | 📡 | **CAN / sensor signals** | **10–100 Hz** |
88
+ | 📁 | **Files per clip** | **10** (1 video + 1 meta + 8 CSV) |
89
+ | 💾 | **Total size** | **~33 GB** |
90
+
91
+ </div>
92
+
93
+ ---
94
+
95
+ ## 🔥 Why ADAS-TO?
96
+
97
+ > *"The takeover moment is the most safety-critical instant in human-automation interaction — yet it remains one of the least studied due to lack of data."*
98
+
99
+ ### 🏆 Unprecedented Scale
100
+ Over **15,000 real-world takeover events** — orders of magnitude larger than existing datasets that typically contain hundreds of events captured in driving simulators.
101
+
102
+ ### 🌍 Unmatched Diversity
103
+ **163 vehicle models** from **23 manufacturers** including Tesla, Toyota, Honda, Hyundai, Ford, Volkswagen, Rivian, and more. From compact EVs to full-size trucks — spanning the full spectrum of modern ADAS implementations.
104
+
105
+ ### 🎯 Rich Multimodal Signals
106
+ Every clip contains **synchronized** front-camera video, vehicle dynamics (speed, acceleration, steering), ADAS controller state, control commands, actuator outputs, driving model predictions, radar/lead vehicle data, and IMU measurements.
107
+
108
+ ### 🌐 Real-World Naturalistic Data
109
+ Collected through **online and offline autonomous driving communities** with diverse real-world driving conditions — highways, urban streets, suburbs, varying weather and lighting. No simulators. No scripted scenarios. Pure naturalistic driving behavior.
110
+
111
+ ---
112
+
113
+ ## 🎯 Use Cases
114
+
115
+ | Application | Description |
116
+ |:---|:---|
117
+ | 🔮 **Takeover Prediction** | Build early warning systems that predict when a driver will need to take over |
118
+ | 🧠 **Driver Behavior Modeling** | Understand human responses during control transitions |
119
+ | 📈 **ADAS Performance Analysis** | Compare disengagement patterns across vehicle types and ADAS systems |
120
+ | 🤖 **Autonomous Driving Safety** | Train and evaluate safety-critical decision-making models |
121
+ | 🧪 **Human Factors Research** | Study cognitive load, reaction times, and situational awareness |
122
+ | 📊 **Multimodal Time-Series** | Develop forecasting and classification models on rich temporal data |
123
+ | 🏗️ **HMI Design** | Design better human-machine interfaces for automated vehicles |
124
+
125
+ ---
126
+
127
+ ## 📁 Dataset Structure
128
+
129
+ ```
130
+ ADAS-TO/
131
+ ├── <CAR_MODEL>/ # e.g., TOYOTA_PRIUS, TESLA_AP3_MODEL_3
132
+ │ └── <driver_XXX>/ # 🔒 anonymized driver ID
133
+ │ └── <route_XXX>/ # 🔒 anonymized route ID
134
+ └── <clip_id>/ # integer (0-indexed per route)
135
+ ── 🎥 takeover.mp4 20-second front-camera video
136
+ ── 📋 meta.json clip metadata & timing
137
+ │ ├── 🚗 carState.csv vehicle dynamics & driver inputs
138
+ │ ├── 🤖 controlsState.csv ADAS controller state & alerts
139
+ │ ├── 🎮 carControl.csv lateral/longitudinal commands
140
+ │ ├── ⚙️ carOutput.csv actuator outputs
141
+ │ ├── 🧠 drivingModelData.csv model predictions & lane detection
142
+ │ ├── 📡 radarState.csv lead vehicle radar data
143
+ │ ├── 📐 accelerometer.csv IMU acceleration data
144
+ ── 📏 longitudinalPlan.csv planner targets & FCW
145
+ ── ...
146
+ ```
147
+
148
+ ---
149
+
150
+ ## 📐 Takeover Event Definition
151
+
152
+ <div align="center">
153
+
154
+ ```
155
+ ◄──────── 10 seconds ────────►◄──────── 10 seconds ────────►
156
+ ┌──────────────────────────────┬──────────────────────────────┐
157
+ │ 🤖 ADAS ENGAGED │ 👤 MANUAL CONTROL │
158
+ │ (automation driving) │ (driver takes over) │
159
+ └──────────────────────────────┴──────────────────────────────┘
160
+
161
+ TAKEOVER EVENT
162
+ (ON → OFF transition)
163
+ ```
164
+
165
+ </div>
166
+
167
+ A **takeover event** is detected as an ADAS ON → OFF transition satisfying:
168
+
169
+ | Criterion | Value |
170
+ |:---|:---|
171
+ | **ADAS engaged** | `controlsState.enabled` OR `cruiseState.enabled` |
172
+ | **Min ON duration** | ≥ 2 seconds before disengagement |
173
+ | **Min OFF duration** | 2 seconds after disengagement |
174
+ | **Gap merging** | Transient gaps < 0.5s merged (filters sensor noise) |
175
+ | **Clip window** | ±10 seconds centered on transition (20s total) |
176
+
177
+ ---
178
+
179
+ ## 📑 Data Fields Reference
180
+
181
+ ### 📋 meta.json Clip Metadata
182
+
183
+ | Field | Type | Description |
184
+ |:---|:---|:---|
185
+ | `car_model` | string | Vehicle model (e.g., `TOYOTA_PRIUS`) |
186
+ | `dongle_id` | string | Anonymized driver ID (`driver_XXX`) |
187
+ | `route_id` | string | Anonymized route ID (`route_XXX`) |
188
+ | `log_kind` | string | Log resolution: `qlog` (10 Hz) or `rlog` (100 Hz) |
189
+ | `log_hz` | int | CAN signal sampling rate |
190
+ | `vid_kind` | string | Camera source type |
191
+ | `camera_fps` | int | Video frame rate (20 fps) |
192
+ | `clip_id` | int | Clip index within route (0-indexed) |
193
+ | `event_mono` | int | Monotonic timestamp of takeover (ns) |
194
+ | `video_time_s` | float | Takeover time within full route video (s) |
195
+ | `clip_start_s` | float | Clip start time within route (s) |
196
+ | `clip_dur_s` | float | Clip duration (s) |
197
+
198
+ ### 🚗 carState.csv Vehicle Dynamics & Driver Inputs
199
+
200
+ | Column | Unit | Description |
201
+ |:---|:---|:---|
202
+ | `vEgo` | m/s | Ego vehicle speed |
203
+ | `aEgo` | m/s² | Ego vehicle acceleration |
204
+ | `steeringAngleDeg` | deg | Steering wheel angle |
205
+ | `steeringTorque` | m | Driver steering torque |
206
+ | `steeringPressed` | bool | Driver actively steering |
207
+ | `gasPressed` | bool | Gas pedal pressed |
208
+ | `brakePressed` | bool | Brake pedal pressed |
209
+ | `cruiseState.enabled` | bool | Cruise / ADAS engaged |
210
+
211
+ ### 🤖 controlsState.csv ADAS Controller
212
+
213
+ | Column | Unit | Description |
214
+ |:---|:---|:---|
215
+ | `enabled` | bool | ADAS system enabled |
216
+ | `active` | bool | ADAS actively controlling vehicle |
217
+ | `curvature` | 1/m | Current path curvature |
218
+ | `desiredCurvature` | 1/m | Target curvature from planner |
219
+ | `vCruise` | m/s | Set cruise speed |
220
+ | `longControlState` | enum | Longitudinal control state |
221
+ | `alertText1` | string | Primary driver alert |
222
+ | `alertText2` | string | Secondary driver alert |
223
+
224
+ ### 🎮 carControl.csv Control Commands
225
+
226
+ | Column | Unit | Description |
227
+ |:---|:---|:---|
228
+ | `latActive` | bool | Lateral control active |
229
+ | `longActive` | bool | Longitudinal control active |
230
+ | `actuators.accel` | m/s² | Commanded acceleration |
231
+ | `actuators.torque` | N·m | Commanded steering torque |
232
+ | `actuators.curvature` | 1/m | Commanded path curvature |
233
+
234
+ ### ⚙️ carOutput.csv Actuator Outputs
235
+
236
+ | Column | Description |
237
+ |:---|:---|
238
+ | `actuatorsOutput.accel` | Acceleration actuator output |
239
+ | `actuatorsOutput.brake` | Brake actuator output |
240
+ | `actuatorsOutput.gas` | Gas actuator output |
241
+ | `actuatorsOutput.steer` | Steering actuator output |
242
+ | `actuatorsOutput.steerOutputCan` | Raw CAN steering output |
243
+ | `actuatorsOutput.steeringAngleDeg` | Steering angle output (deg) |
244
+
245
+ ### 🧠 drivingModelData.csv Driving Model Predictions
246
+
247
+ | Column | Description |
248
+ |:---|:---|
249
+ | `action.desiredCurvature` | Model-predicted desired curvature |
250
+ | `action.desiredAcceleration` | Model-predicted desired acceleration |
251
+ | `laneLineMeta.leftProb` | Left lane line detection probability |
252
+ | `laneLineMeta.rightProb` | Right lane line detection probability |
253
+
254
+ ### 📡 radarState.csv Lead Vehicle Detection
255
+
256
+ | Column | Unit | Description |
257
+ |:---|:---|:---|
258
+ | `leadOne.dRel` | m | Distance to primary lead vehicle |
259
+ | `leadOne.vRel` | m/s | Relative velocity of lead |
260
+ | `leadOne.vLead` | m/s | Absolute velocity of lead |
261
+ | `leadOne.aLeadK` | m/s² | Lead vehicle acceleration |
262
+ | `leadTwo.*` | | Secondary lead vehicle (same fields) |
263
+
264
+ ### 📐 accelerometer.csvIMU Data
265
+
266
+ | Column | Unit | Description |
267
+ |:---|:---|:---|
268
+ | `acceleration.v` | m/s² | 3-axis acceleration vector |
269
+ | `timestamp` || Sensor timestamp |
270
+
271
+ ### 📏 longitudinalPlan.csvPlanner Outputs
272
+
273
+ | Column | Unit | Description |
274
+ |:---|:---|:---|
275
+ | `aTarget` | m/s² | Target acceleration |
276
+ | `hasLead` | bool | Lead vehicle detected |
277
+ | `fcw` | bool | Forward collision warning active |
278
+ | `speeds[]` | m/s | Planned speed profile |
279
+ | `accels[]` | m/s² | Planned acceleration profile |
280
+
281
+ ---
282
+
283
+ ## 🚘 Vehicle Coverage
284
+
285
+ <div align="center">
286
+
287
+ **23 Manufacturers · 163 Models · From Compact EVs to Full-Size Trucks**
288
+
289
+ </div>
290
+
291
+ ### Top Vehicle Models by Clip Count
292
+
293
+ | # | Vehicle Model | Clips | | # | Vehicle Model | Clips |
294
+ |:---:|:---|---:|:---:|:---:|:---|---:|
295
+ | 1 | 🏆 RIVIAN R1 GEN1 | 2,127 | | 10 | CHEVROLET BOLT EUV | 244 |
296
+ | 2 | 🥈 ACURA MDX 3G | 1,863 | | 11 | TOYOTA RAV4 TSS2 | 228 |
297
+ | 3 | 🥉 FORD F-150 MK14 | 1,226 | | 12 | RAM HD 5TH GEN | 221 |
298
+ | 4 | CHEVROLET SILVERADO | 639 | | 13 | VOLKSWAGEN JETTA MK7 | 215 |
299
+ | 5 | TOYOTA PRIUS | 482 | | 14 | KIA EV6 | 209 |
300
+ | 6 | HONDA CIVIC | 470 | | 15 | VOLKSWAGEN GOLF MK7 | 192 |
301
+ | 7 | TESLA MODEL 3 | 432 | | 16 | KIA NIRO EV | 185 |
302
+ | 8 | FORD MAVERICK MK1 | 300 | | 17 | HYUNDAI IONIQ 6 | 177 |
303
+ | 9 | HYUNDAI IONIQ 5 | 266 | | 18 | VOLKSWAGEN ATLAS MK1 | 153 |
304
+
305
+ <details>
306
+ <summary>📋 <b>All 23 Manufacturers</b> (click to expand)</summary>
307
+
308
+ > Acura · Audi · BYD · Chevrolet · Ford · Genesis · Honda · Hyundai · Jeep · Kia · Lexus · Mazda · Nissan · Porsche · RAM · Rivian · Skoda · Subaru · Tesla · Toyota · Volkswagen · Volvo
309
+
310
+ </details>
311
+
312
+ ---
313
+
314
+ ## 🚀 Quick Start
315
+
316
+ ### Loading a Single Clip
317
+
318
+ ```python
319
+ import json
320
+ import pandas as pd
321
+ from huggingface_hub import hf_hub_download
322
+
323
+ repo_id = "HenryYHW/ADAS-TO"
324
+ clip_path = "TOYOTA_PRIUS/driver_001/route_001/0"
325
+
326
+ # 📋 Download metadata
327
+ meta_path = hf_hub_download(repo_id, f"{clip_path}/meta.json", repo_type="dataset")
328
+ with open(meta_path) as f:
329
+ meta = json.load(f)
330
+
331
+ # 🚗 Load vehicle state signals
332
+ car_state = pd.read_csv(
333
+ hf_hub_download(repo_id, f"{clip_path}/carState.csv", repo_type="dataset")
334
+ )
335
+ print(car_state[["vEgo", "aEgo", "steeringAngleDeg", "brakePressed"]].describe())
336
+
337
+ # 🤖 Load ADAS controller state
338
+ controls = pd.read_csv(
339
+ hf_hub_download(repo_id, f"{clip_path}/controlsState.csv", repo_type="dataset")
340
+ )
341
+
342
+ # 📡 Load radar data
343
+ radar = pd.read_csv(
344
+ hf_hub_download(repo_id, f"{clip_path}/radarState.csv", repo_type="dataset")
345
+ )
346
+ ```
347
+
348
+ ### Iterating Over All Clips
349
+
350
+ ```python
351
+ from huggingface_hub import HfApi
352
+
353
+ api = HfApi()
354
+ files = api.list_repo_files("HenryYHW/ADAS-TO", repo_type="dataset")
355
+ meta_files = [f for f in files if f.endswith("meta.json")]
356
+ print(f"Total clips: {len(meta_files)}") # → 15,705
357
+ ```
358
+
359
+ ### 💾 Download the Full Dataset
360
+
361
+ ```bash
362
+ # Using huggingface-cli (recommended)
363
+ huggingface-cli download HenryYHW/ADAS-TO --repo-type dataset --local-dir ./ADAS-TO
364
+
365
+ # Using git-lfs
366
+ git lfs install
367
+ git clone https://huggingface.co/datasets/HenryYHW/ADAS-TO
368
+ ```
369
+
370
+ ---
371
+
372
+ ## 🔒 Privacy & Ethics
373
+
374
+ - **Anonymized identifiers**: All driver and route IDs are replaced with anonymous tokens (`driver_XXX`, `route_XXX`)
375
+ - **Forward-view only**: Video captures road-facing view only — no cabin or driver footage
376
+ - **No PII**: No personally identifiable information is included in any data file
377
+ - **Community-sourced**: Data collected through autonomous driving enthusiast communities with informed participation
378
+
379
+ ---
380
+
381
+ ## 📖 Data Collection
382
+
383
+ ADAS-TO was built from naturalistic driving logs contributed by **online and offline autonomous driving communities**. Participating drivers voluntarily shared their driving data collected through various ADAS-equipped vehicles during everyday driving. The raw logs were processed through an automated pipeline to:
384
+
385
+ 1. **Detect** ADAS disengagement events (ON→OFF transitions)
386
+ 2. **Extract** synchronized video and CAN-bus signals within a 20-second window
387
+ 3. **Validate** each clip for signal completeness and temporal alignment
388
+ 4. **Anonymize** all driver and route identifiers
389
+
390
+ This community-driven collection approach enables unprecedented scale and diversity, capturing genuine driver behavior across a wide spectrum of vehicles, road types, and driving conditions.
391
+
392
+ ---
393
+
394
+ ## 📝 Citation
395
+
396
+ If you use ADAS-TO in your research, please cite:
397
+
398
+ ```bibtex
399
+ @dataset{adas_to_2026,
400
+ title = {ADAS-TO: A Large-Scale Multimodal Naturalistic Dataset and
401
+ Empirical Characterization of Human Takeovers during ADAS Engagement},
402
+ author = {Anonymous Authors},
403
+ year = {2026},
404
+ publisher = {Hugging Face},
405
+ url = {https://huggingface.co/datasets/HenryYHW/ADAS-TO}
406
+ }
407
+ ```
408
+
409
+ ---
410
+
411
+ ## 📄 License
412
+
413
+ <div align="center">
414
+
415
+ This dataset is released under [**CC BY-NC 4.0**](https://creativecommons.org/licenses/by-nc/4.0/).
416
+
417
+ For academic and non-commercial research purposes.
418
+
419
+ ---
420
+
421
+ *Built with ❤️ for the autonomous driving research community*
422
+
423
+ </div>