4D Radar Road Boundaries and Freespace

What It Is

This page covers 4D radar methods for road-boundary detection and radar-derived drivable freespace.
The anchor method is 4DRadarRBD, a 2025 4D mmWave radar road-boundary curve detector.
4DRadarRBD uses radar point clouds to detect static physical edges of the available driving area, such as fences, bushes, and roadblocks.
The broader freespace connection is that boundary curves and radar occupancy maps both define where the vehicle should not drive.
This is separate from radar object detection and from dense 4D radar-camera occupancy.
For dense radar occupancy, see 4D Radar-Camera Occupancy.

4D mmWave radar provides range, azimuth, elevation, and Doppler, enabling all-weather geometric sensing.
Road boundaries reflect millimeter waves, producing radar points that can be segmented from clutter.
4DRadarRBD first filters noisy points using physical constraints for plausible road boundaries.
It then performs point-wise boundary segmentation with a distance-based loss that penalizes false boundary detections far from true boundaries.
It adds temporal dynamics by comparing each point to the motion-compensated road-boundary estimate from the previous frame.
It fits continuous boundary curves from segmented points so planning can consume curves rather than disconnected returns.

Input: 4D mmWave radar point cloud.
Per-point radar features: x, y, z, Doppler velocity, signal-to-noise ratio, and range.
Vehicle-state input: velocity and yaw rate from GPS/IMU or the localization stack.
Optional input for freespace systems: radar occupancy probabilities or BEV grid outputs.
Output: point-wise road-boundary segmentation.
Output: continuous left/right or multi-curve road-boundary estimates.
Planning output: drivable corridor, freespace boundary, or boundary constraints for path planning.

Module 1: preprocess radar points, extract point-wise features, reduce noise, and fuse frames to mitigate sparsity.
Module 2: segment boundary points using a PointNet++-style point-cloud network.
Add a distance-based loss so far-away false positives are penalized more strongly.
Add temporal-deviation features from the previous motion-compensated boundary estimate.
Module 3: cluster detected boundary points and fit continuous road-boundary curves.
For freespace use, convert boundary curves or radar occupancy maps into drivable-space polygons with confidence.
Keep radar ghost and multipath signals visible in diagnostics rather than silently smoothing them away.

4DRadarRBD is evaluated through real-world driving tests.
The Frontiers paper reports 93% road-boundary point segmentation accuracy.
It reports a median distance or Chamfer error of 0.023 m and a 92.6% error reduction versus the baseline model.
On twisting mountain-road cases, the paper reports 91.2% point-wise segmentation accuracy and a median Chamfer distance of 0.11 m.
Evaluation should include boundary point accuracy, curve error, temporal stability, false boundary creation, and missed-boundary gaps under occlusion.
Freespace evaluation should use both free-space precision and collision-relevant false-free metrics.

Radar is robust to darkness, glare, fog, rain, and some spray conditions where cameras degrade.
4D elevation helps distinguish overhead structures from boundaries that block motion.
Boundary curves provide a compact planner-friendly representation.
Temporal features improve stability despite radar sparsity and ego motion.
Radar freespace can complement camera lane/curb perception and LiDAR geometry.
The method is valuable where object taxonomy is incomplete but the drivable corridor still matters.

Radar multipath and ghost reflections can create false boundaries near metal, glass, fences, aircraft, or terminal structures.
Large occluding vehicles can block returns from true boundaries and split curves.
Low radar point density can miss thin cones, ropes, chocks, tow bars, hoses, or low curbs.
Boundary-style training may not generalize to open aprons where drivable areas are marked by paint or procedures rather than physical barriers.
Temporal propagation can carry a false boundary forward if confidence is not checked.
A detected boundary is not a complete freespace proof; interior obstacles still need detection or occupancy.

Strong adverse-weather fit for service roads, stand edges, terminal curbs, fence lines, and fixed barriers.
Useful at night and in fog, rain, jet exhaust, or de-icing mist where camera-only freespace is weak.
Airport aprons often lack raised road boundaries, so the method must be adapted to cones, painted stand limits, aircraft no-go zones, and mapped geofences.
Radar around aircraft can be noisy because fuselage, engines, landing gear, and terminal glass create strong reflections.
Use radar boundary output as one layer in a drivable-space stack with maps, LiDAR, cameras, and operations rules.
Never treat radar freespace as empty near aircraft unless small-object/FOD coverage is separately validated.

Preserve raw radar features, SNR, Doppler, elevation, and per-point timestamps.
Use ego-motion compensation from the production localization stack, not a loosely synchronized estimate.
Log boundary confidence and temporal carry-over source for every fitted curve segment.
Add airport-specific false-positive tests around aircraft stands, jet bridges, glass facades, wet pavement, and metallic GSE.
Fuse with HD map no-go zones so radar cannot open forbidden areas.
Validate downstream planning behavior with missing-boundary, false-boundary, and boundary-jump scenarios.