V2X-Fusion: Multi-Modal BEV Perception with V2X Message Integration

In an Intelligent Transportation System (ITS), collaborative perception plays a vital role in enhancing road safety. Most collaborative perception studies utilize images and point clouds for fusion decision-making. This may cause low sensing precision under Non-Line-of-Sight (NLOS) conditions. To address this issue, we propose a unified V2X-Fusion framework that leverages Vehicle-to-Everything (V2X) messages to augment environmental perception. In particular, we propose two multi-modal cooperative perception modes: a lightweight model for real-time efficiency, and a comprehensive model for higher accuracy. These methods integrate image, point cloud, and V2X message to enhance road perception in complex traffic scenarios. In the fusion module, we construct bird's-eye view (BEV) representations for each modality. We employ a masking mechanism to distinguish between known and inferred regions to avoid semantic conflicts during fusion. We conducted extensive experiments on the nuScenes dataset. The results show that our methods significantly improve BEV perception accuracy, particularly in cases of occlusion and long-range targets.