Vehicle–infrastructure cooperative (VIC) perception systems enhance ego-centric 3D detection by leveraging complementary measurements from infrastructure sensors, alleviating issues such as limited field of view and occlusions in vehicle-mounted devices. Existing methods predominantly use homogeneous architectures, sharing the same input types and dense feature representations between vehicles and roadside units, which leads to high computational and communication costs.
To address these limitations, we propose HeteroCoopDet, a heterogeneous cooperative perception framework. Infrastructure sensors fuse LiDAR and camera data to produce rich multimodal features, while the vehicle side uses a lightweight vision-dominant pipeline to reduce redundancy. A Cross-Agent Interaction (CAI) module efficiently fuses object-query-level features between agents, enabling effective integration of complementary information.
Experiments on the DAIR-V2X dataset show that Hetero-CoopDet outperforms current homogeneous cooperative models, achieving an mAP@0.3 of 0.8199. These results validate that our heterogeneous design, combined with cross-agent query interaction, significantly improves detection performance while mitigating the overheads of conventional VIC systems.