HighwayScene

HighwayScene is a synchronized multi-LiDAR dataset recorded from a static roadside installation at a highway construction site. It was introduced for the cross-sensor benchmark in Beam-Wise Statistical Background Subtraction for Static Roadside LiDAR: A Cross-Sensor Benchmark Study.

The recording contains free-flowing, dense highway traffic under a legally enforced speed limit of 40 km/h. Three LiDAR technologies observe the same scene:

Sensor	Principle	Nominal rate
Ouster OS0	Rotating time-of-flight	10 Hz
Aeva Aeries II	FMCW	10 Hz
Blickfeld QB2	Solid-state time-of-flight	5 Hz

Project resources:

• Project page
• Background-subtraction code
• Protobuf reader
• Dataset repository

Dataset structure

The original protobuf records are published without generated caches or model outputs. Record filenames encode their inclusive global frame-ID range and must not be renamed.

HighwayScene/
├── metadata/
│   ├── ground_truth.yaml
│   ├── manifest.json
│   └── SHA256SUMS
├── train/
│   └── *.pb
├── val/
│   └── *.pb
└── test/
    └── *.pb

Split	Record files	Global frame IDs	Records	Ouster frames	Aeva frames	Blickfeld frames
train	20	1–4000	4000	4000	4000	2008
validation	5	4001–5000	1000	1000	1000	503
test	5	5001–5998	998	998	994	503
total	30	1–5998	5998	5998	5994	3014

Sensor-frame counts differ because the Blickfeld sensor operates at 5 Hz and some records do not contain a usable point cloud for every sensor.

File format and loading

Each .pb file contains protobuf-serialized frames with the available LiDAR point clouds, sensor identifiers, scan indices, calibration values, and Aeva radial velocities. The reference implementation uses highwayscene-proto==1.0.0.

python -m pip install highwayscene-proto==1.0.0

Download a local snapshot with huggingface_hub:

from huggingface_hub import snapshot_download

dataset_root = snapshot_download(
    repo_id="iis-esslingen/HighwayScene",
    repo_type="dataset",
    local_dir="/path/to/HighwayScene",
)

Read the first frame from a downloaded record:

from pathlib import Path

from highwayscene import decode_pointcloud, iter_frames

record_path = next((Path(dataset_root) / "test").glob("*.pb"))
frame = next(iter_frames(record_path))
points = decode_pointcloud(frame.lidars[0].pointcloud)
print(frame.frame_id, points.xyz.shape)

The reference adapter and complete paper configurations are provided by the code repository. After installing the project, a paper reproduction run can be started with:

static-bg-subtraction reproduce-paper \
  --highwayscene-root /path/to/HighwayScene \
  --coopscenes-root /path/to/CoopScenes \
  --output-dir outputs/paper

Ground truth

HighwayScene uses two reproducible static/dynamic labeling strategies:

• Ouster and Blickfeld points are labeled from fixed lane-aligned 3D regions of interest and the corresponding ground-plane definitions.
• Aeva points are labeled dynamic when the absolute measured radial velocity exceeds 1 m/s.

The immutable parameters used for the paper benchmark are stored in metadata/ground_truth.yaml and are mirrored by the tagged reference code. The raw dataset does not contain separate generated point-wise label files.

Integrity

metadata/manifest.json describes the released split and sensor counts. metadata/SHA256SUMS contains a SHA-256 checksum for every protobuf record and the ground-truth snapshot.

Intended use and limitations

The dataset supports research on background modeling, dynamic-point segmentation, roadside LiDAR processing, and comparisons across heterogeneous scan patterns. The benchmark methods assume a statically mounted sensor with a known, fixed scan pattern and a temporally stable static background. Results may not transfer directly to moving sensors, changed sensor poses, or scenes with substantial long-term structural changes.

License

HighwayScene is released under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International license. See LICENSE for the complete legal text.

Suggested attribution: HighwayScene, Alexander Baumann, Marcel Voßhans, and Thao Dang, CC BY-NC-SA 4.0. Please also cite the accompanying paper.

Commercial use is not permitted under this license. For questions about commercial use, contact alexander.baumann@hs-esslingen.de.

Citation

If you use HighwayScene, please cite the accompanying paper:

@inproceedings{baumann2026beamwise,
  author    = {Alexander Baumann and Marcel Vo{\ss}hans and Thao Dang},
  title     = {Beam-Wise Statistical Background Subtraction for Static
               Roadside {LiDAR}: A Cross-Sensor Benchmark Study},
  booktitle = {IEEE International Conference on Intelligent Transportation
               Systems (ITSC)},
  year      = {2026}
}

Final DOI and proceedings metadata will be added after publication.

Authors

• Alexander Baumann
• Marcel Voßhans
• Thao Dang

Institute for Intelligent Systems, Esslingen University of Applied Sciences, Germany. Contact: alexander.baumann@hs-esslingen.de.