DenoiseCP-Net

What It Is

DenoiseCP-Net is a 2025 LiDAR-based collective perception method for adverse weather.
It jointly performs voxel-level denoising and 3D object detection for cooperative vehicles.
The key motivation is that adverse-weather noise hurts perception and also wastes communication bandwidth when noisy LiDAR features are shared.
It addresses rain, snow, and fog in a simulated OPV2V extension.
It is not a standalone single-vehicle denoiser like LiSnowNet, TripleMixer, or AdverseNet; it is tied to cooperative perception.

Denoise before communication so cooperative vehicles do not transmit non-informative weather noise.
Integrate voxel-level noise filtering and object detection into a unified sparse convolution backbone.
Avoid a two-stage pipeline where denoising and detection run separate backbones over the same LiDAR data.
Remove weather-corrupted voxels while preserving detection accuracy.
Reduce inference latency, compute, and collective-perception bandwidth together.
Extend OPV2V with realistic rain, snow, and fog simulations to study collective perception under adverse weather.

Input: voxelized LiDAR point clouds from ego and cooperative vehicles.
Input: cooperative perception data shared among vehicles or agents.
Training input: OPV2V scenes augmented with simulated rain, snow, and fog.
Intermediate output: shared sparse-convolution features used for both denoising and detection.
Output: voxel-level noise filtering decisions.
Output: 3D object detections for collective perception.
Non-output: no camera/radar fusion, no real airport communication protocol, and no direct model for road spray, de-icing mist, dust, steam, or multipath.

Simulate adverse weather on OPV2V LiDAR scenes for rain, snow, and fog.
Voxelize noisy LiDAR observations from cooperative agents.
Process the voxel grid with a sparse convolution backbone.
Attach denoising and object-detection heads to the shared computation path.
Remove or suppress weather-noise voxels before communication or downstream fusion.
Perform collective perception with reduced transmitted noise and maintained detection accuracy.
Evaluate detection, denoising accuracy, bandwidth, and latency as coupled system metrics.

The paper claims the first study of LiDAR-based collective perception under diverse adverse weather conditions.
It evaluates on an OPV2V extension with simulated rain, snow, and fog.
The arXiv abstract reports near-perfect denoising accuracy in adverse weather.
It reports bandwidth reduction up to 23.6 percent while maintaining the same detection accuracy.
It also reports reduced inference latency for cooperative vehicles by eliminating redundant two-stage computation.
Evidence is simulation-centered and should not be read as proof of real-world production readiness.

Addresses a system-level issue that single-vehicle denoisers miss: noisy weather points consume V2X bandwidth and latency budget.
Joint denoising and detection avoids redundant feature extraction.
Sparse convolution is a practical representation for voxel-level cooperative perception.
Evaluation couples denoising with object detection, bandwidth, and latency.
Rain, snow, and fog are all in scope.
The approach fits connected autonomy better than a filter that only cleans the ego cloud after communication.

OPV2V is simulated; the reported weather extension may not capture airport de-icing mist, road spray, steam, dust, wet concrete reflections, or multipath ghosts.
Cooperative perception depends on synchronization, calibration, pose accuracy, packet loss, and trust between agents.
Removing voxels before communication can delete rare safety-critical evidence if the denoising head is wrong.
Fog and heavy precipitation can cause missing returns, not just extra noisy voxels.
Dynamic objects seen by one agent and not another can be confused with noise without robust temporal/fusion logic.
The arXiv paper does not establish a production communication standard or safety case.
Real deployment needs cybersecurity, V2X scheduling, latency bounds, and fallback behavior beyond the denoising model.

Strong conceptual fit for airports with connected AVs, smart infrastructure, or cooperative ramp vehicles.
Bandwidth-aware denoising matters when multiple vehicles share LiDAR around gates, service roads, and aprons.
Needs real airside V2X data, not only OPV2V simulation, before drawing operational conclusions.
Must include non-road artifacts: aircraft, GSE, jet bridges, cones, chocks, personnel, reflective surfaces, de-icing clouds, and service-vehicle spray.
Pair with Radar-LiDAR Fusion in Adverse Weather because radar may preserve object evidence when LiDAR denoising is uncertain.
Use Production Perception Systems for deployment constraints around monitoring, fallbacks, and auditability.

Preserve both pre-denoising and post-denoising voxel sets in logs so communication savings can be audited against missed detections.
Track denoising mistakes by object class, range, weather type, and agent viewpoint.
Validate the method under packet loss, timestamp jitter, pose error, and mismatched LiDAR models.
Do not tune solely for bandwidth reduction; enforce minimum detection recall for small and low-profile hazards.
Add airside weather labels for snow, rain, fog, road spray, de-icing mist/steam, dust, ghost/multipath, and dynamic-object disagreement.
Compare against running AdverseNet or TripleMixer before communication as separate baselines.
Treat simulation gains as a hypothesis until verified with real cooperative logs.