BiVIBEV: Bimodal Vehicle-Infrastructure Cooperative Perception via Unified BEV Representation

Infrastructure sensors provide broader perception ranges and fewer occlusions, making vehicle-infrastructure cooperative perception increasingly vital. Existing methods mostly adopt LiDAR point clouds from both vehicles and infrastructure to generate Bird's Eye View (BEV) representations, followed by adaptive fusion. Despite their effectiveness, these LiDAR-based approaches fundamentally neglect the semantic richness and textural details inherent in camera information, which limits the deployment capability of vehicle-infrastructure cooperative systems in practical scenarios. For vehicle-infrastructure cooperative 3D object detection, BiVIBEV is proposed, a novel framework. This work designs the Vehicle-Infrastructure Spatial Cross-Attention (VI-SCA) module to acquire unified image BEV features for subsequent fusion. This module utilizes LiDAR BEV representations from both vehicles and infrastructure as geometric priors, engaging in iterative interactions with the corresponding vehicle-view and infrastructure-view images. Moreover, we incorporate an occupancy supervision branch designed to provide denser geometric-level supervision for BEV features. Experimental results on the DAIR-V2X-C dataset demonstrate that BiVIBEV surpasses existing methods and achieves state-of-the-art (SOTA) performance.