K-Radar

What It Is

K-Radar is the KAIST-Radar 4D radar dataset and benchmark for autonomous driving.
It provides raw 4D radar tensor data rather than only pre-filtered radar point clouds.
The dataset is designed for object detection in diverse weather and road conditions.
It includes synchronized auxiliary sensors for calibration, labeling, and fusion research.
The official repository provides dataset tools, annotation tools, visualization, and baseline code.
It is one of the core public references for 4D radar perception.

Preserve full radar information across range, azimuth, elevation, and Doppler.
Provide 4D radar tensors with power measurements instead of collapsing early to sparse point detections.
Annotate 3D boxes so radar-native object detectors can be trained and benchmarked.
Include difficult weather where radar should outperform optical sensors.
Enable comparisons between 4D radar, LiDAR, camera, and fusion baselines.
Demonstrate why elevation is important for 3D object detection from radar.

Dataset input: 4D radar tensor data with Doppler, range, azimuth, and elevation dimensions.
Dataset input: calibrated high-resolution LiDAR.
Dataset input: surround stereo camera data.
Dataset input: IMU/RTK-GPS or pose-related auxiliary measurements.
Labels: carefully annotated 3D bounding boxes for road objects.
Benchmark output: trained detector predictions and standard 3D object detection metrics.

The dataset contains about 35K frames.
Radar measurements are represented as 4DRT, or 4D radar tensor, in full RAED form.
Conditions include fog, rain, snow, and other challenging weather.
Road structures include urban, suburban, alleyway, and highway settings.
The repository includes GUI tools for annotation, visualization, calibration, and inference inspection.
Baseline neural networks are included for 4DRT-based object detection.

The paper trains baseline neural networks on 4D radar tensors.
It compares radar-based and similarly structured LiDAR-based networks under adverse weather.
The authors show height/elevation information is crucial for 3D object detection.
Evaluation focuses on 3D box detection from radar and sensor-fusion settings.
Weather-conditioned analysis is one of the dataset's main values.
The dataset should be treated as road-weather evidence, not direct airport evidence.

Full RAED radar tensor supports research beyond sparse radar point-cloud detection.
Adverse-weather coverage is directly relevant to robust autonomy.
Synchronized LiDAR and camera support multimodal fusion experiments.
Public tools make the dataset practical for radar detector development.
35K frames provide more scale than many earlier radar datasets.
Useful for studying when radar remains stable while LiDAR or cameras degrade.

Road scenes do not include aircraft, jet bridges, or ramp-specific clutter.
Radar hardware characteristics may differ from production sensors selected for an airside vehicle.
Tensor data is heavier than point clouds and may require specialized preprocessing.
3D boxes alone do not cover free space, semantics, or small FOD detection.
Multipath near large metallic structures is underrepresented compared with airports.
Weather diversity does not include de-icing spray, jet blast, glycol mist, or ramp floodlight glare.

Strong evidence base for making 4D radar a primary perception input in adverse weather.
Good starting dataset for radar detector pretraining before airside fine-tuning.
Supports evaluating radar-LiDAR fusion when LiDAR density drops in rain, fog, or snow.
Needs airport-specific data collection because aircraft surfaces and terminal infrastructure change radar clutter.
Radar tensor access is valuable if production hardware allows low-level data export.
Airside safety cases should cite K-Radar as supporting evidence, not as validation coverage.

Use K-Radar to evaluate candidate radar encoders before collecting airport data.
Decide early whether the target radar exposes tensors or only point clouds; this changes model choice.
Build conversion scripts to a common BEV coordinate frame shared with LiDAR.
Compare tensor-based models against RadarPillars-style point-cloud models.
Add weather metadata and point-count/tensor-energy diagnostics to validation reports.
Plan for radar-specific annotation review because boxes can be visible in radar when LiDAR is weak.