IEEE ITSC 2026 · Open paper and dataset

Beam-Wise Statistical Background Subtraction for Static Roadside LiDAR: A Cross-Sensor Benchmark Study

Learning the static world beam by beam—without semantic training labels—and benchmarking it across rotating, solid-state, and FMCW LiDAR.

Alexander Baumann · Marcel Vosshans · Thao Dang

Institute for Intelligent Systems, Esslingen University of Applied Sciences

Read the paper Explore the dataset View the code

Synchronized HighwayScene point clouds from rotating time-of-flight, FMCW, and solid-state LiDAR. — One roadside scene, observed by three LiDAR technologies. Static background Dynamic foreground

3LiDAR technologies

5,998HighwayScene records

4background models

5benchmark scenes

The work

A practical benchmark for static roadside LiDAR

Statistical models learn the expected range of every beam from a short, unlabeled sequence and flag points that do not match the static scene.

Model every beam

Four lightweight model families capture the range statistics of a fixed scan pattern—without semantic annotation or a pretrained network.

Compare across sensors

The same evaluation spans rotating time-of-flight, solid-state time-of-flight, and FMCW LiDAR as well as four urban CoopScenes sequences.

Reuse the benchmark

Data, labels, a typed reader, final configurations, and modular method and dataset interfaces support reproduction and follow-up work.

HighwayScene dataset

Same traffic. Same viewpoint. Different scan geometry.

A synchronized multi-LiDAR recording from a static roadside installation at a highway construction site, released with reproducible train, validation, and test splits.

Open HighwayScene on Hugging Face

Annotated Ouster OS0 point cloud. — Rotating time-of-flight **Ouster OS0** 10 Hz

Annotated AEVA Aeries II point cloud. — FMCW LiDAR **AEVA Aeries II** 10 Hz

Annotated Blickfeld QB2 point cloud. — Solid-state time-of-flight **Blickfeld QB2** 5 Hz

4,000 train 1,000 validation 998 test 100 m benchmark range

Benchmark result

High accuracy without a semantic model

Beam-wise GMM reaches the highest filtered F1 score for all three HighwayScene sensors. Density-aware CFTA is the most consistent method across the complete cross-dataset benchmark.

Ouster OS0

99.43%

Filtered F1 · GMM

Blickfeld QB2

96.95%

Filtered F1 · GMM

AEVA Aeries II

97.59%

Filtered F1 · GMM

Qualitative HighwayScene FMCW background-subtraction result. — **HighwayScene** FMCW result on dense highway traffic

Qualitative CoopScenes background-subtraction result. — **CoopScenes** Transfer to an urban cooperative scene

Best filtered scores on the HighwayScene test split. Full precision, recall, and cross-scene results are reported in the paper.

See the full benchmark

Open resources

Continue with the paper, data, or code

Paper, data, labels, and implementation resources for the benchmark.

Paper Read on arXiv Methods, experiments, and complete results Dataset HighwayScene Records, splits, metadata, and documentation Code Reproduce and extend Python implementation and paper configurations Labels CoopScenes benchmark Manually created 3D annotations for four sequences

BibTeX citation

@misc{baumann2026beamwise,
  author        = {Alexander Baumann and Marcel Vosshans and Thao Dang},
  title         = {Beam-Wise Statistical Background Subtraction for Static Roadside {LiDAR}: A Cross-Sensor Benchmark Study},
  year          = {2026},
  eprint        = {2608.14868},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CV},
  url           = {https://arxiv.org/abs/2608.14868},
  note          = {Accepted at IEEE ITSC 2026}
}