README.md

---
license: cc-by-nc-sa-4.0
task_categories:
  - other
tags:
  - fiftyone
  - multimodal
  - lidar
  - roadside-perception
  - background-subtraction
  - intelligent-transportation-systems
  - point-cloud
  - mcap
pretty_name: HighwayScene
size_categories:
  - n<1K
dataset_summary: |
  HighwayScene is a synchronized multi-LiDAR roadside dataset recorded at a highway construction site. Three statically mounted sensors — an Ouster OS0 (rotating ToF, 10 Hz), an Aeva Aeries II (FMCW, 10 Hz), and a Blickfeld QB2 (solid-state ToF, 5 Hz) — observe the same free-flowing highway traffic scene simultaneously. This FiftyOne version re-encodes the original protobuf records as MCAP episodes for exploration in the FiftyOne multimodal viewer.

  The dataset contains 30 episodes (~20 seconds each), split into train (20), val (5), and test (5). Each episode carries three synchronized point-cloud streams with per-point fields including radial velocity (Aeva) and beam identifiers (Ouster, Blickfeld).

  ## Installation

  ```bash
  pip install -U fiftyone
  ```

  ## Usage

  ```python
  import fiftyone as fo
  from fiftyone.utils.huggingface import load_from_hub

  dataset = load_from_hub("Voxel51/HighwayScene")
  session = fo.launch_app(dataset)
  ```
---

# Dataset Card for HighwayScene
![image/png](highwayscene.gif)

## Installation

```bash
pip install -U fiftyone
```

## Usage

```python
import fiftyone as fo
from fiftyone.utils.huggingface import load_from_hub

dataset = load_from_hub("Voxel51/HighwayScene")
session = fo.launch_app(dataset)
```

## Dataset Details

### Dataset Description

HighwayScene is a synchronized multi-LiDAR roadside dataset introduced for the cross-sensor background subtraction benchmark in "Beam-Wise Statistical Background Subtraction for Static Roadside LiDAR: A Cross-Sensor Benchmark Study" (ITSC 2026). Three statically mounted sensors with different sensing principles observe the same highway construction-site scene simultaneously:

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

The scene features consistently high traffic volume under a legally enforced 40 km/h speed limit. The dataset supports research on background modeling, dynamic-point segmentation, and cross-sensor benchmarking for static roadside LiDAR.

This FiftyOne version converts the original protobuf `.pb` records to MCAP format so each episode can be explored interactively in the FiftyOne multimodal viewer, with all three sensor streams on a shared 20-second timeline.

- **Curated by:** Alexander Baumann, Marcel Voßhans, Thao Dang — Institute for Intelligent Systems, Esslingen University of Applied Sciences, Germany
- **Funded by:** [More Information Needed]
- **Shared by:** Institute for Intelligent Systems (IIS), Esslingen University of Applied Sciences
- **Language(s):** N/A
- **License:** CC BY-NC-SA 4.0 (non-commercial use only)

### Dataset Sources

- **Repository:** https://huggingface.co/datasets/iis-esslingen/HighwayScene (original protobuf records)
- **Paper:** https://arxiv.org/abs/2608.14868
- **Demo:** https://highwayscene.github.io/

## Uses

### Direct Use

- Benchmarking beam-wise statistical background subtraction methods across heterogeneous LiDAR technologies
- Cross-sensor evaluation of static/dynamic point-cloud segmentation
- Research on roadside LiDAR perception: traffic monitoring, infrastructure-based sensing, inter-vehicle distance estimation
- Multi-modal LiDAR data exploration via the FiftyOne App's 3D tile and shared timeline

### Out-of-Scope Use

- Moving-sensor or ego-vehicle applications — the benchmark assumes a statically mounted sensor with a fixed scan pattern
- Commercial use — the CC BY-NC-SA 4.0 license prohibits commercial applications
- Scenes with substantial long-term structural changes, changed sensor poses, or dynamically varying scan patterns

## Dataset Structure

### FiftyOne Dataset Topology

**Media type:** `multimodal` (MCAP files)
**Sample count:** 30 episodes
**Splits:** train (20), val (5), test (5), encoded as the `split` string field and queryable via `dataset.match(fo.ViewField("split") == "test")`

Each FiftyOne sample points to one `.mcap` file covering approximately 20 seconds of synchronized multi-sensor recording. The MCAP file contains three point-cloud topics (`/ouster/points`, `/aeva/points`, `/blickfeld/points`) and three static identity frame transforms on `/tf_static`. All topics use the `foxglove.PointCloud` schema and are rendered in the FiftyOne App's **3D tile** on a shared timeline.

### Sample Fields

| Field | FiftyOne type | Description |
|-------|---------------|-------------|
| `filepath` | `StringField` | Absolute path to the `.mcap` episode file |
| `split` | `StringField` | Benchmark split: `"train"`, `"val"`, or `"test"` |
| `record_name` | `StringField` | Original `.pb` filename stem (encodes recording date and global frame-ID range) |
| `first_frame_id` | `IntField` | First global frame ID in this episode (1-based, dataset-wide) |
| `last_frame_id` | `IntField` | Last global frame ID in this episode |
| `duration_s` | `FloatField` | Episode duration in seconds (~19.9 s for 200-frame episodes) |
| `message_count` | `IntField` | Total MCAP message count across all topics in this episode |
| `ouster_frames` | `IntField` | Number of Ouster OS0 point-cloud messages (10 Hz; 200 per full episode) |
| `aeva_frames` | `IntField` | Number of Aeva Aeries II point-cloud messages (10 Hz; 194–200 per episode) |
| `blickfeld_frames` | `IntField` | Number of Blickfeld QB2 point-cloud messages (5 Hz; 95–101 per episode) |
| `topics` | `ListField(StringField)` | MCAP topic names present in the episode |
| `schemas` | `ListField(StringField)` | MCAP schema names used (`foxglove.PointCloud`, `foxglove.FrameTransform`) |
| `has_pointcloud` | `BooleanField` | `True` for all 30 episodes |
| `has_image` | `BooleanField` | `False` — no camera data in this dataset |
| `has_gps` | `BooleanField` | `False` — no GPS data |
| `has_imu` | `BooleanField` | `False` — no IMU data |
| `has_logs` | `BooleanField` | `False` — no log messages |

### MCAP Topic Structure (per episode)

| Topic | Schema | Rate | Per-point fields |
|-------|--------|------|-----------------|
| `/ouster/points` | `foxglove.PointCloud` | 10 Hz | `x`, `y`, `z` (float32); `intensity`, `ambient`, `reflectivity` (float32); `channel_id` (float32) |
| `/aeva/points` | `foxglove.PointCloud` | 10 Hz | `x`, `y`, `z` (float32); `intensity`, `ambient`, `reflectivity`, `velocity` (float32, radial m/s) |
| `/blickfeld/points` | `foxglove.PointCloud` | 5 Hz | `x`, `y`, `z` (float32); `reflectivity`, `channel_id`, `horizontal_id` (float32) |
| `/tf_static` | `foxglove.FrameTransform` | static | Identity transforms for `ouster`, `aeva`, `blickfeld` frames relative to `world` |

Typical point counts per frame: Ouster ~45K, Aeva ~92K, Blickfeld ~12K. Blickfeld has no `intensity` or `ambient` channel. Aeva's `velocity` field (radial velocity in m/s) is the unique FMCW capability enabling velocity-based dynamic-point labeling.

### Ground Truth

No per-frame label files are stored in the FiftyOne dataset. The benchmark computes static/dynamic labels at evaluation time from parameters in `metadata/ground_truth.yaml` (in the original repository):
- **Ouster and Blickfeld**: lane-aligned 3D bounding-box ROIs define dynamic regions; any point inside is labeled dynamic.
- **Aeva**: a point is labeled dynamic if its measured radial velocity exceeds 1 m/s.

### Parsing Decisions

- Original records are protobuf `.pb` files decoded with `highwayscene-proto==1.0.0`. Each `.pb` → one `.mcap` using `foxglove-sdk==0.26.0`.
- All uint16 auxiliary channels (intensity, ambient, reflectivity, channel_id, horizontal_id) are cast to float32 in the MCAP point-cloud buffers.
- Points with range ≤ 0.5 m are filtered (zero-return guard).
- Static transforms use `foxglove.FrameTransform` with no embedded timestamp, placing them in the viewer's static store for the full episode duration.
- No extrinsic sensor-to-sensor calibration is provided by the dataset authors; all `world → sensor` transforms are identity.
- Blickfeld operates at 5 Hz: frames where the sensor is absent are simply omitted from the `/blickfeld/points` topic.

## Dataset Creation

### Curation Rationale

Existing roadside LiDAR datasets focus on object detection or semantic segmentation with class-based labels, and are typically limited to a single sensor or scene. HighwayScene was created to fill the gap in systematic cross-sensor evaluation for static background subtraction — providing simultaneous recordings from three heterogeneous LiDAR technologies under real highway traffic conditions, with reproducible annotation strategies that do not require manual point-wise labeling.

### Source Data

#### Data Collection and Processing

Data was recorded at a German highway construction site with a legally enforced 40 km/h speed limit. Sensors were mounted statically on roadside infrastructure. Recording rate is 10 Hz for the Ouster and Aeva sensors and 5 Hz for the Blickfeld. All streams are temporally synchronized. 5,998 global frames are captured across the full recording session (Feb 9, 2026), partitioned into train (frames 1–4000), val (4001–5000), and test (5001–5998) splits using a strict non-overlapping protocol.

#### Who are the source data producers?

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

### Annotations

#### Annotation process

Static/dynamic ground-truth labels are not stored per frame in the dataset records. Instead, two reproducible annotation strategies are defined in `metadata/ground_truth.yaml`:

1. **Geometry-based (Ouster, Blickfeld):** For each traffic lane, a static 3D bounding-box ROI aligned with the roadway is manually defined. Under empty-scene conditions no LiDAR returns fall inside these volumes. Since vehicles are constrained by concrete barriers, any return inside a lane volume originates from dynamic traffic. Points inside are labeled dynamic; all others static.

2. **Velocity-based (Aeva):** A point is labeled dynamic if its measured FMCW radial velocity exceeds 1 m/s. This strategy requires no scene geometry and illustrates the FMCW sensor's self-annotating capability.

#### Who are the annotators?

Annotation parameters (lane ROIs, ground planes) were defined by the dataset authors. No manual point-wise labeling was performed — labels are reproducibly derived from geometry and physics.

#### Personal and Sensitive Information

LiDAR point clouds do not record visually identifiable information (faces, license plates). No personal data is present in this dataset.

## Citation

**BibTeX:**

```bibtex
@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}
}
```

**APA:**

Baumann, A., Voßhans, M., & Dang, T. (2026). Beam-Wise Statistical Background Subtraction for Static Roadside LiDAR: A Cross-Sensor Benchmark Study. *IEEE International Conference on Intelligent Transportation Systems (ITSC)*.

## More Information

- Original dataset (protobuf format): https://huggingface.co/datasets/iis-esslingen/HighwayScene
- Background subtraction code: https://github.com/HighwayScene/roadside-lidar-background-subtraction
- Protobuf reader: https://github.com/HighwayScene/highwayscene-proto
- Project page: https://highwayscene.github.io/

## Dataset Card Authors

Harpreet Sahota

## Dataset Card Contact

harpreetsahota07@gmail.com