RCFusionNet: A Radar-Guided Two-Stage Fusion Approach to Effective Multimodal Perception

Multisensor fusion perception plays a pivotal role in advancing autonomous driving. While most existing LiDAR–camera (LC) fusion methods achieve high detection accuracy, LiDAR is inherently vulnerable to extreme weather conditions and is costlier compared to microwave radar. In this work, we introduce a robust radar–camera (RC) fusion model, RCFusionNet, aimed at delivering enhanced and robust detection performances under adverse weather conditions. Specifically, we introduce a camera feature encoder (CFE) to enhance the capability of visual representations, which incorporate multiscale bird’s-eye view (BEV) features to further enhance the detection performance. Additionally, a radar feature encoder (RFE) is introduced to fully exploit radar point cloud data by selectively extracting informative radar features. Furthermore, we adopt a radar-guided two-stage fusion strategy (RGTSFS), applied during both the BEV feature generation and the RC feature fusion phases. Extensive evaluations on the nuScenes benchmark demonstrate the effectiveness of our RCFusionNet.