4D Radar Meets LiDAR and Camera: Cooperative Perception under Adverse Weather
Melih Yazgan, Iramm Hamdard, Qiyuan Wu, J. Marius Zoellner · 2026
Read 4D Radar Meets LiDAR and Camera: Cooperative Perception under Adverse Weather by Melih Yazgan, Iramm Hamdard, Qiyuan Wu, J. Marius Zoellner (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2606-00416.txt.
Introduction
Yazgan and co-authors add 4D radar to team sight for self-drive cars in bad weather. Cams and laser sight fade in fog and rain. Team sight, where cars share what they see, turns frail then. Radar holds up in such weather and adds speed traits as a fourth strand.
Their blend uses radar speed cues to guide how shared views merge. Two backbones are stretched to take radar in the mix. In one, radar stands in for faded laser sight. In the next, it backs laser sight up. New bench sets pair radar with past sets, with fake laser wear built in. Fog and rain tests gain much from radar in both forms. Truck set tests show the gain holds past sims.
Catalog record
| CREATOR | Melih Yazgan, Iramm Hamdard, Qiyuan Wu, J. Marius Zoellner |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 34 KB |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2606.00416
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
4D Radar Meets LiDAR and Camera: Cooperative Perception under Adverse Weather Melih Yazgan, Iramm Hamdard, Qiyuan Wu, J. Marius Zoellner Abstract Cooperative perception is important for autonomous driving but remains fragile when cameras and LiDAR degrade in adverse weather. We address this challenge by integrating 4D imaging radar as a weather-robust modality into collaborative perception and introducing a Doppler-guided spatial attention mechanism for multi-agent fusion. Our approach extends two representative backbones: a radar-camera pipeline where radar substitutes LiDAR, and a LiDAR-radar pipeline where radar complements LiDAR. To support evaluation, we release radar-augmented benchmarks, OPV2V-R and Adver-City-R, with physics-based LiDAR degradation. Experiments show strong robustness gains in fog and rain, including substantial improvements when radar replaces degraded LiDAR. Additional validation on MAN TruckScenes demonstrates transfer beyond simulation. Overall, our results highlight 4D imaging radar as a robust modality for all-weather collaborative perception. Dataset and code are available at: https://url.fzi.de/SlimComm . 1 Introduction The safe deployment of autonomous vehicles hinges on reliable perception. While autonomous driving has made remarkable progress in recent years, particularly in environmental perception 16 ; 33 ; 18 , failures in perception directly translate into safety risks. Robust perception remains a prerequisite for large-scale deployment 26 , yet adverse weather such as rain and fog can severely degrade camera- and LiDAR-based systems 32 . Cameras suffer from reduced visibility, while LiDAR suffers from scattering and attenuation, leading to degraded detection and tracking performance. To mitigate these vulnerabilities, modern autonomous platforms adopt multi-sensor fusion strategies that exploit complementary strengths across modalities 32 ; 37 ; 6 . Among them, four-dimensional (4D) imaging radar has attracted growing interest because it remains robust in poor visibility and directly provides Doppler velocity measurements 24 ; 45 . Its resilience makes 4D radar a promising complementary modality for all-weather perception, yet it remains underexplored in collaborative frameworks. Figure 1 : Impact of adverse weather on LiDAR and radar perception. Top: LiDAR point clouds degrade heavily under fog and rain, losing density and range. Bottom: Radar remains stable, with Doppler cues highlighting movers. This complementary robustness motivates our radar-augmented collaborative fusion. (Simulated via CARLA 2 and the LISA 12 framework). A parallel trend is collaborative perception, where connected vehicles and infrastructure share sensor information to overcome occlusion and extend perceptual range. Recent surveys and benchmarks highlight its potential and open challenges 8 ; 9 ; 40 . However, the majority of collaborative perception methods have been evaluated predominantly under clear-weather conditions, with limited systematic treatment of multi-modal sensor degradation under adverse weather 42 . LiDAR, the dominant modality in collaborative frameworks, is also fragile in adverse weather, as illustrated in Fig. 1 . Progress is further bottlenecked by the absence of suitable datasets. Widely used resources such as OPV2V 35 , Adver-City 11 , and DAIR-V2X 44 each miss essential requirements such as synchronized 4D radar, realistic weather degradation, or full 360° coverage. Reviews of collaborative-perception datasets confirm that no public benchmark currently satisfies the combined need for 4D radar, 360° layouts, and physics-based weather simulation 41 ; 31 . SlimComm 43 previously integrated radar into OPV2V and Adver-City and benchmarked LiDAR-radar collaborative perception, but left weather-induced sensor degradation unaddressed. Our contributions are threefold: 1. Radar-augmented collaborative fusion: We integrate 4D radar into two representative collaborative baselines: substituting LiDAR in BM2CP to form a radar-camera pipeline, and complementing LiDAR in Where2comm to form a LiDAR-radar pipeline. 2. Doppler-guided spatial attention: We introduce an attention mechanism that leverages radar Doppler velocity to create a dynamic Bird’s-Eye-View (BEV) mask, effectively emphasizing moving objects across multi-agent features. 3. Adverse-weather benchmarks and evaluation: We release OPV2V-R and Adver-City-R with physics-based weather degradation. We demonstrate strong robustness against severe weather, communication latency, and spatial misalignment, and validate our modules on the real-world MAN TruckScenes 3 single-vehicle dataset to confirm transferability beyond simulation. 2 Related Work 2.1 Radar Fusion in Single-Vehicle Perception Autonomous vehicles typically rely on cameras, LiDAR, and radar as core exteroceptive sensors. Cameras provide rich semantics but degrade under poor visibility, while LiDAR delivers precise geometry but suffers from scattering and attenuation in rain and fog 45 ; 27 . Radar, by contrast, is inherently resilient to visibility degradation and directly provides Doppler velocity cues. The emergence of 4D radar has therefore motivated a range of fusion methods that exploit radar either in combination with cameras, LiDAR, or as a standalone modality. Early approaches such as CRFNet 21 and RAMP-CNN 4 integrated radar detections with images, while CenterFusion 20 , RCFusion 47 , and CRAFT 13 demonstrated improved detection under challenging conditions. More recent works leverage 4D radar directly: InterFusion 30 proposed interaction-based LiDAR-radar fusion, LXL 34 excluded LiDAR entirely with a radar occupancy-assisted depth strategy, and MSSF 17 introduced a multi-stage radar-camera sampling framework. RadarPillars 19 further advanced radar-only detection by enriching velocity features with intra-pillar offsets and applying radar-tailored attention. Taken together, these approaches confirm 4D radar as a weather-resilient and cost-effective modality for perception. However, they remain confined to the single-vehicle setting and do not address multi-agent collaboration. 2.2 Collaborative Perception under Adverse Weather Collaborative perception extends perceptual range and mitigates occlusions by allowing connected vehicles to exchange features. Most prior works have focused on LiDAR-camera pipelines 28 ; 10 ; 15 , showing that collaboration improves robustness but still inherits LiDAR’s fragility in adverse weather. Jiang et al. 10 addressed uncertainty in rain via probabilistic feature aggregation, while Li et al. 15 proposed domain generalization to unseen weather. Tsakmakopoulou and Moustakas 28 showed that V2V communication helps mitigate fog degradation in cooperative pipelines by integrating their proposed method on top of S-AdaFusion 25 . Complementing these, Wang et al. 29 surveyed adverse-weather cooperative strategies, emphasizing the need for efficiency under communication constraints. Taking a different approach, Huang et al. 9 introduced AttFuse w/MDD, a denoising diffusion model that uses weather-robust 4D radar features to guide the reconstruction of noisy LiDAR features. While advanced techniques like denoising diffusion show promise for improving accuracy, they can also introduce significant computational overhead. In contrast, our work prioritizes an efficient solution, and no prior collaborative perception method has explicitly leveraged radar Doppler cues to guide attention across agents in a lightweight manner. 2.3 Datasets for Collaborative Radar Perception Progress in collaborative perception relies heavily on dedicated datasets. Large-scale single-vehicle datasets such as nuScenes 1 , K-Radar 22 , and MAN TruckScenes 3 include radar and adverse-weather conditions but lack cooperative settings. Collaborative datasets such as OPV2V 35 , DAIR-V2X 44 , and Adver-City 11 support multi-agent fusion, but OPV2V lacks realistic weather, DAIR-V2X has restricted modality coverage, and Adver-City does not model LiDAR degradation. TUMTraf-V2X 48 provides real-world V2X cooperative perception with camera and LiDAR, but does not include radar as a cooperative modality. Radar-focused cooperative datasets remain limited. V2X-R 9 provides simulated radar perception but without full 360° coverage and with Doppler encoded only as a boolean flag. V2X-Radar 39 introduces real-world radar data, but its public release omits cooperative elements and suffers from fidelity issues in ego-velocity and Doppler. SlimComm 43 contributed radar-augmented versions of OPV2V and Adver-City, enabling early LiDAR-radar cooperative benchmarks, but without modeling adverse weather. Summary. Existing works confirm the promise of radar-camera fusion for single-vehicle perception, with specialized designs such as