HighwayScene · Static Roadside LiDAR

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

Alexander Baumann¹ Marcel Voßhans¹ Thao Dang¹

¹ Institute for Intelligent Systems, Esslingen University of Applied Sciences, Germany

Accepted at the 29th IEEE International Conference on Intelligent Transportation Systems (ITSC 2026)

arXiv coming soon Code coming soon HighwayScene data CoopScenes labels coming soon

Synchronized HighwayScene observations from three roadside LiDAR technologies. — Synchronized HighwayScene observations from the Ouster OS0, AEVA Aeries II, and Blickfeld QB2. Dynamic foreground points are shown in red.

3LiDAR technologies

5,998HighwayScene records

5evaluated scenes

>90%best filtered F1 per scene

Overview

Background subtraction without semantic training labels

We study statistical background subtraction for statically mounted roadside LiDAR. The methods learn the range behavior of each beam from an unlabeled recording and identify points that are inconsistent with the learned static background during inference.

The benchmark combines HighwayScene—a synchronized recording from solid-state, rotating time-of-flight, and FMCW LiDAR—with four urban CoopScenes sequences. It compares four beam-wise model families and spatial consistency filters under a shared evaluation protocol.

New multi-LiDAR dataset

A synchronized static highway scene recorded with three complementary sensing technologies.

Four model families

Density-aware CFTA, quantile-constrained Gaussian, beam-wise GMM, and image-based MOG2.

Reproducible benchmark

Released configurations, reference results, dataset adapters, and extensible method interfaces.

Method

Beam-wise modeling at a glance

The approach assumes a fixed, known scan pattern and a temporally stable static background. Training requires only a short unlabeled point-cloud sequence.

1 Observe Collect 600 static-sensor frames.

2 Model Fit independent range statistics per beam.

3 Predict Classify points against the learned background.

4 Filter Enforce spatial consistency in angular or 3D space.

Dataset

HighwayScene

A synchronized multi-LiDAR recording from a static roadside installation at a highway construction site.

Annotated point cloud from the Blickfeld QB2 solid-state LiDAR. — **Blickfeld QB2**
Solid-state LiDAR · 5 Hz

Annotated point cloud from the Ouster OS0 rotating time-of-flight LiDAR. — **Ouster OS0**
Rotating time-of-flight LiDAR · 10 Hz

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

One scene, three sensing principles

All sensors observe the same traffic scene from a fixed roadside position. The published split contains 4,000 training, 1,000 validation, and 998 test records. Dynamic references are derived from lane-aligned 3D regions for the Ouster and Blickfeld streams and radial velocity for the FMCW stream.

Recording rate: 10 Hz; Blickfeld 5 Hz
Paper range: 100 m
Traffic setting: Highway construction zone
Evaluation: Point-wise static/dynamic labels

Results

Strong performance across sensor technologies

Spatial filtering substantially improves precision. On HighwayScene, beam-wise GMM achieves the highest filtered F1 score for all three sensors; density-aware CFTA is the most consistent method across the full 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

Values are the best filtered F1 scores on the HighwayScene test split. The benchmark additionally evaluates transfer to four urban CoopScenes sequences.