The dataset viewer is not available for this split.
Error code: StreamingRowsError
Exception: ValueError
Message: Dataset 'range_candidate_bins' has length 13 but expected 201
Traceback: Traceback (most recent call last):
File "/src/services/worker/src/worker/utils.py", line 149, in get_rows_or_raise
return get_rows(
dataset=dataset,
...<4 lines>...
column_names=column_names,
)
File "/src/libs/libcommon/src/libcommon/utils.py", line 272, in decorator
return func(*args, **kwargs)
File "/src/services/worker/src/worker/utils.py", line 129, in get_rows
rows_plus_one = list(itertools.islice(safe_iter(ds, dataset=dataset), rows_max_number + 1))
File "/src/services/worker/src/worker/utils.py", line 544, in safe_iter
yield from ds.decode(False) if ds.features else ds
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 3052, in __iter__
for key, example in ex_iterable:
^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2559, in __iter__
for key, pa_table in self._iter_arrow():
~~~~~~~~~~~~~~~~^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2584, in _iter_arrow
for key, pa_table in self.ex_iterable._iter_arrow():
~~~~~~~~~~~~~~~~~~~~~~~~~~~~^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 558, in _iter_arrow
for key, pa_table in iterator:
^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 440, in _iter_arrow
for key, pa_table in self.generate_tables_fn(**gen_kwags):
~~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/hdf5/hdf5.py", line 76, in _generate_tables
num_rows = _check_dataset_lengths(h5, self.info.features)
File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/hdf5/hdf5.py", line 353, in _check_dataset_lengths
raise ValueError(f"Dataset '{path}' has length {dset.shape[0]} but expected {num_rows}")
ValueError: Dataset 'range_candidate_bins' has length 13 but expected 201Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.
- Overview
- Session Attributes
- HDF5 File Structure
- Dataset Organisation and Method-Specific Data Files
- Radar Data
- Physiological Signals
- Vision Signals
- PVT Labels
- Motion Scores
- Temporal Alignment
- Processed Features
- Loading the Dataset
- Intended Use
- Code and Implementation
- Citation
- Data Availability and Access
- License
ICR-DS: In-Cabin Radar Dataset
ICR-DS (In-Cabin Radar Dataset) is a multimodal naturalistic driving dataset for driver vigilance and physiological state monitoring. The dataset was collected from 28 participants during approximately 50 minutes of real-world urban and highway driving using synchronised wearable physiological sensing, radar, and an in-cabin RGB camera.
To provide an objective measure of vigilance, each participant completed a Psychomotor Vigilance Test (PVT) before and after each drive. The dataset therefore supports longitudinal analysis of behavioural and physiological driver-state measurements, as well as the development of non-contact physiological monitoring using in-cabin radar sensing.
Overview
Each recording session is stored as a single HDF5 file. The file contains:
- Raw signals retained at their respective acquisition or processing resolutions.
- Processed physiological and vision metrics computed through temporal aggregation.
- Radar tensors generated from the raw radar measurements.
- Session-level attributes describing the participant, timestamps, sensor configuration, and modality availability.
- PVT summaries collected before and after the drive.
- Derived motion scores aligned with the radar recording timeline.
The dataset includes synchronised recordings from:
- A 60 GHz FMCW radar sensor.
- A BioHarness 3 chest strap.
- An in-cabin RGB camera.
- VBOX-recorded vehicle data.
- Session-level PVT outcomes.
OpenFace was applied to the video stream to derive eye, gaze, and head-behaviour measurements.
Session Attributes
Each HDF5 file contains session-level attributes describing the recording.
| Attribute | Example | Description |
|---|---|---|
session_id |
"s027_2026-05-08" |
Session identifier |
subject_id |
"s027" |
Participant identifier |
start_time_iso |
"2026-05-08T11:27:21Z" |
Session start time in UTC ISO-8601 format |
end_time_iso |
"2026-05-08T11:28:32Z" |
Session end time in UTC ISO-8601 format |
frame_period_s |
0.09 |
Radar frame period in seconds |
carrier_freq_hz |
60000000000 |
Radar carrier frequency |
wavelength_m |
0.005 |
Radar wavelength |
range_resolution_m |
0.0843 |
Radar range resolution |
has_radar |
"present" |
Radar availability flag |
has_vision |
"present" |
Vision availability flag |
has_bioharness |
"present" |
BioHarness availability flag |
has_gps |
"present" |
GPS availability flag |
has_pvt |
"present" |
PVT availability flag |
HDF5 File Structure
Each session file follows the structure below:
/
├── raw/
│ ├── radar/
│ │ ├── beamformed_cube
│ │ └── localization_cube
│ │
│ ├── physiology/
│ │ ├── hr_1hz
│ │ ├── br_1hz
│ │ ├── ibi
│ │ └── ibi_ts
│ │
│ ├── vision/
│ │ ├── landmarks_raw
│ │ ├── gaze_xy
│ │ ├── pose_rxyz
│ │ ├── eye_open
│ │ └── confidence
│ │
│ └── labels/
│ ├── predrive
│ ├── postdrive
│ ├── body_motion_score
│ └── vehicle_motion_score
│
└── processed/
├── physiology/
│ ├── mean_hr
│ ├── mean_br
│ ├── rmssd
│ ├── sdnn
│ ├── pnn50
│ └── hf
│
└── vision/
├── blink_rate
├── perclos
├── blink_duration
├── gaze_speed
├── gaze_x_std
├── gaze_y_std
├── head_speed
├── head_yaw_std
├── head_pitch_std
└── head_roll_std
The exact availability of individual datasets may vary depending on the availability of the corresponding sensor recordings.
Dataset Organisation and Method-Specific Data Files
The dataset is organised into training and testing partitions.
Within each partition, the STAT/ directory contains the main
session-level HDF5 files, which store the synchronised radar,
physiological, vision and label data described in this documentation.
The STAT directory also contains the corresponding PyTorch (.pt)
files for direct use with PyTorch workflows.
The remaining directories contain method-specific HDF5 files prepared for the corresponding baseline heart-rate estimation approaches.
The directory structure is organised as follows:
ICR-DS/
├── Train/
│ ├── STAT/
│ │ ├── *.h5
│ │ └── *.pt
│ ├── HRConvNet/
│ │ └── *.h5
│ ├── Alizadeh/
│ │ └── *.h5
│ ├── VitaNet/
│ │ └── *.h5
│ └── mmFormer/
│ └── *.h5
│
└── Test/
├── STAT/
│ ├── *.h5
│ └── *.pt
├── HRConvNet/
│ └── *.h5
├── Alizadeh/
│ └── *.h5
├── VitaNet/
│ └── *.h5
└── mmFormer/
└── *.h5
STAT
STAT is provided in two file formats:
- HDF5 (
.h5): Contains the processed radar data, including the beamformed and localisation representations used in STAT preprocessing, together with associated timing information. - PyTorch (
.pt): Contains the model-ready STAT input tensors, aligned reference heart-rate labels and associated window information, suitable for direct use with PyTorch.
Both formats are provided to support different data-loading workflows.
Other methods
The following methods are provided in HDF5 format:
- HRConvNet: Eight-mode empirical wavelet transform (EWT)-based representations derived from radar slow-time signals in the selected range region.
- Alizadeh: Radar slow-time signals from the strongest moving range bin selected within the 13-bin range region.
- VitaNet: Phase-derived radar time-series representations across 13 range bins and 12 virtual antenna channels.
- mmFormer: Complex I/Q radar data from four receiver channels and four selected range bins (7–10).
Each method-specific HDF5 file contains the radar input representation prepared for the corresponding method, together with associated window timing. Reference physiological measurements are linked to the corresponding session-level data where applicable.
The files are organised according to the training and testing partitions used throughout the dataset.
Two participants did not have BioHarness recordings and were therefore excluded from these partitions, which require reference physiological measurements. Their recordings are retained separately outside the Train/ and Test/ directories.
For details of the preprocessing procedures, data preparation scripts and model-specific implementations, please refer to the accompanying GitHub repository.
Radar Data
Radar measurements are stored in two synchronised branches under:
raw/radar/beamformed_cuberaw/radar/localization_cube
Both datasets contain the same radar windows, with matching temporal indices and tensor dimensions.
Radar tensor dimensions
Each radar window is stored as a real-valued tensor containing the real and imaginary components of the complex beamformed radar signal.
The tensor shape is:
(Window, Frame, IQ, Doppler, Range, Angle)
For the current preprocessing configuration:
(N, 200, 2, 9, 13, 32)
| Dimension | Size | Description |
|---|---|---|
| Window | N | Number of consecutive radar windows in the recording |
| Frame | 200 | Radar frames within each window |
| IQ | 2 | Real and imaginary components of the complex signal |
| Doppler | 9 | Selected Doppler bins centred on zero Doppler |
| Range | 13 | Selected range bins, corresponding to range bins 3–15 |
| Angle | 32 | Azimuth angle FFT bins |
The radar frame period is 90 ms, corresponding to approximately 11.11 Hz.
Each 200-frame radar window therefore represents approximately 18 seconds of observation.
Range-bin selection
The radar processing retains 13 range bins (indices 3–15 inclusive) to cover the region expected to contain the driver's chest.
Based on the in-vehicle cabin geometry and the expected distance between the radar and the driver, the chest is most likely to lie within range bins 7–10. The wider 3–15 region is retained to provide coverage around this expected location rather than restricting the radar data to a narrow range interval.
Individual heart-rate estimation methods may use this range region differently, depending on their respective input requirements.
Phase-preserving radar branch
raw/radar/beamformed_cube contains the phase-preserving radar representation.
This branch is generated from the range FFT without subtracting the chirp/slow-time mean. The coherent quasi-static radar return is retained to preserve the inter-frame phase variations required for phase-based physiological signal extraction.
Localisation radar branch
raw/radar/localization_cube contains a separate clutter-suppressed representation used for chest localisation.
It is generated by subtracting the chirp/slow-time mean from a separate copy of the range FFT before Doppler and azimuth processing.
The original phase-preserving branch is not modified during this operation.
Both branches use the same radar processing configuration, range selection, Doppler selection, and azimuth processing. They are stored in lock-step so that window index n refers to the same recording interval in both datasets.
The localisation branch supports fixed chest-region selection without modifying the coherent signal used for subsequent physiological analysis.
STAT input derivation
For STAT, the localisation branch is used to identify the driver's chest region and guide spatial selection. The selected chest range bin and strongest azimuth bin are identified, with the two neighbouring angle bins also retained.
The corresponding phase-preserving radar signals are extracted using five Doppler bins centred on zero Doppler. This produces 15 complex slow-time channels, comprising one range bin, three angle bins and five Doppler bins (1 × 3 × 5).
These signals are subsequently processed into spectral features to form the STAT model input. The resulting model-ready inputs, together with aligned reference heart-rate labels and window information, are provided in the STAT PyTorch (.pt) files.
Further details of the chest localisation, angular selection, Doppler selection and spectral processing are provided in the accompanying paper.
Physiological Signals
The BioHarness 3 chest strap provided reference physiological measurements, including heart rate, breathing rate and beat-to-beat timing information.
| Measurement | Temporal Resolution | Unit | Description |
|---|---|---|---|
| Heart rate | 1 Hz | bpm | Mean heart rate within each one-second interval. |
| Breathing rate | 1 Hz | breaths/min | Mean breathing rate within each one-second interval. |
| Interbeat intervals (IBIs) | Event-based | ms | Time intervals between successive heartbeats. |
| Beat timings | Event-based | s | Timestamps of individual heartbeat events, represented as elapsed time relative to the start of the recording. |
Stored in:
raw/physiology/hr_1hz
raw/physiology/br_1hz
raw/physiology/ibi
raw/physiology/ibi_ts
These measurements provide physiological references for heart-rate and breathing-rate analysis and support the derivation of heart-rate variability (HRV) features from the IBI data.
Vision Signals
The in-cabin RGB camera was mounted next to the radar and focused on the driver's head and face.
The video was acquired at approximately 30 frames per second. OpenFace was applied to the video stream to extract facial landmark, gaze, and head-pose measurements on a frame-by-frame basis, resulting in an approximate feature sampling rate of 30 Hz.
| Feature | Temporal Resolution | Unit | Description |
|---|---|---|---|
| Facial landmarks | Approx. 30 Hz | pixels | Two-dimensional x and y coordinates of 12 eye-region facial landmarks (points 36–47). |
| Gaze angles | Approx. 30 Hz | rad | Horizontal and vertical eye-gaze angles. |
| Head pose | Approx. 30 Hz | rad | Head rotation components around the x, y and z axes. |
| Eye aspect ratio | Approx. 30 Hz | - | Eye aspect ratio calculated separately for the left and right eyes and averaged to obtain a frame-level measurement. |
| Confidence | Approx. 30 Hz | - | OpenFace confidence score associated with facial detection. |
Stored in:
raw/vision/landmarks_raw
raw/vision/gaze_xy
raw/vision/pose_rxyz
raw/vision/eye_open
raw/vision/confidence
These measurements support the analysis of eye behaviour and closure, gaze dynamics, and head movement.
PVT Labels
To provide objective measures of vigilance, participants completed a PVT before and after each drive.
The following outcomes were extracted:
| Metric | Unit | Description |
|---|---|---|
mean_rt_ms |
ms | Mean reaction time across the PVT task. |
median_rt_ms |
ms | Median reaction time across the PVT task. |
reciprocal_rt_1_per_s |
s⁻¹ | Reciprocal reaction-time measure, representing response speed. |
lapse_rate |
– | Rate of attentional lapses during the task. |
commission_rate |
– | Rate of commission errors during the task. |
Stored in:
raw/labels/predrive
raw/labels/postdrive
The pre-drive and post-drive measurements support the assessment of vigilance and the investigation of changes across the driving session.
Motion Scores
Two separate motion scores are stored in:
raw/labels/body_motion_score
raw/labels/vehicle_motion_score
These scores provide measures of body movement and vehicle movement during each recording.
Body motion score
The body motion score is derived from BioHarness measurements using two features:
| Feature | Description |
|---|---|
Abody |
Euclidean magnitude of the peak-to-minimum acceleration ranges across the three accelerometer axes |
activity_score |
BioHarness activity measurement |
The original measurements are grouped by UTC second, with available observations averaged within each second.
Vehicle motion score
The vehicle motion score is derived from GPS/VBOX measurements using three features:
| Feature | Description |
|---|---|
accel_mean_abs |
Mean absolute longitudinal acceleration |
jerk_mean_abs |
Mean absolute longitudinal jerk |
VV_std_m_s |
Standard deviation of vertical vehicle velocity within the second |
Vehicle acceleration and jerk are derived from changes in vehicle speed using the actual elapsed time between consecutive GPS samples.
GPS gaps are handled explicitly. Derivatives are not calculated across sampling intervals exceeding the configured gap threshold of 0.5 seconds.
Temporal resolution, normalisation and score calculation
Both motion scores are calculated independently for each 18-second radar window.
A minimum of 15 valid one-second feature observations is required within each 18-second block. Blocks that do not meet this coverage requirement are assigned NaN.
The motion features are normalised using the median and interquartile range calculated from the training cohort. Normalised feature values are clipped to the range [-5, 5].
Each modality score is calculated as the equally weighted mean of its normalised features, provided that all required features are available for that block.
Temporal Alignment
Each modality has its own time stream.
The shared analysis interval is defined using the latest common start time and the earliest end time across the synchronised recordings.
The HDF5 session attributes start_time_iso and end_time_iso define the session-level UTC time interval.
Feature extraction begins after a 5-minute offset to enable reliable HRV estimation and maintain alignment across modalities.
Processed Features
Processed features are computed from the raw signals using fixed temporal aggregation windows.
Physiology Metrics
Physiological features are computed every minute using a 5-minute rolling window.
| Metric | Unit | Description |
|---|---|---|
| Heart rate | bpm | Mean heart rate. |
| Breathing rate | breaths/min | Mean breathing rate. |
| RMSSD | ms | Root mean square of successive differences between adjacent IBIs. |
| SDNN | ms | Standard deviation of IBIs. |
| HF power | ms² | High-frequency spectral power of IBI variability. |
| pNN50 | % | Percentage of successive IBI differences exceeding 50 ms. |
Stored in:
processed/physiology/
Vision Metrics
Blink-based measures are computed over non-overlapping 1-minute windows.
Gaze and head-pose features are computed over non-overlapping 10-second windows to capture short-term changes in visual scanning behaviour.
Blink Metrics
| Metric | Unit | Description |
|---|---|---|
| Blink rate | blinks/min | Number of blink events per minute |
| Blink duration | s | Mean duration of detected blinks |
| PERCLOS | % | Percentage of time the eyes are closed |
Gaze Metrics
| Metric | Unit | Description |
|---|---|---|
| Gaze x std | rad | Variability of horizontal gaze angle |
| Gaze y std | rad | Variability of vertical gaze angle |
| Gaze speed | rad/s | Mean rate of gaze-angle change |
Head Pose Metrics
| Metric | Unit | Description |
|---|---|---|
| Yaw std | rad | Variability of yaw angle |
| Pitch std | rad | Variability of pitch angle |
| Roll std | rad | Variability of roll angle |
| Head speed | rad/s | Mean rate of head-pose change |
Stored in:
processed/vision/
Loading the Dataset
Each HDF5 file corresponds to a single driving session.
A session can be loaded using h5py:
import h5py
with h5py.File("s006_session.h5", "r") as f:
# Radar
radar_phase = f["raw/radar/beamformed_cube"]
radar_localization = f["raw/radar/localization_cube"]
# Physiology
physiology = f["raw/physiology"]
# Vision
vision = f["raw/vision"]
# Labels
labels = f["raw/labels"]
body_motion = labels["body_motion_score"]
vehicle_motion = labels["vehicle_motion_score"]
# Processed features
processed_physiology = f["processed/physiology"]
processed_vision = f["processed/vision"]
The radar datasets can be accessed by window index:
# Read the first 18-second phase-preserving radar window
phase_window = radar_phase[0]
# Read the corresponding localisation window
localization_window = radar_localization[0]
# Read the corresponding motion scores
body_score = body_motion[0]
vehicle_score = vehicle_motion[0]
Intended Use
ICR-DS is intended to support research on:
- Development and benchmarking of non-contact, radar-based heart-rate estimation methods for in-vehicle environments.
- Investigation of the effects of driver body movement and vehicle dynamics on radar-based heart-rate estimation performance.
- Multimodal driver-state analysis combining radar, vision, physiological reference measurements, motion scores and vigilance outcomes.
- Investigation of relationships between driver visual behaviour, physiological responses and vigilance.
- Comparison of pre-drive and post-drive vigilance outcomes to examine changes across driving sessions.
Code and Implementation
The accompanying GitHub repository provides the code associated with this dataset and the corresponding research experiments.
It includes the data organisation and preparation scripts, radar data generation procedures, model implementations, training scripts and evaluation workflows.
For implementation details and instructions for reproducing the experiments, please refer to the ICR-DS GitHub repository.
Citation
If you use this dataset in your research, please cite the associated publication. Citation details will be provided when the publication is available.
Data Availability and Access
Raw video recordings: Original in-cabin RGB video recordings are not included in the public dataset release due to ethical and participant privacy requirements.
Raw radar ADC data: Original radar ADC recordings are not included in the public release. Researchers interested in accessing these recordings for research purposes are welcome to contact the dataset authors.
For questions about ICR-DS or raw radar data requests, please contact:
Email: anima.rahman.1@warwick.ac.uk
License
This dataset is licensed under the Creative Commons Attribution 4.0 International (CC BY 4.0) license.
For the full license text, see:
- Downloads last month
- 426