Radar-vision fusion offers a cost-effective solution for autonomous driving perception, combining robust radar ranging with rich camera semantics. However, the significant modality gap between sparse radar Bird’s-Eye-View (BEV) geometry and dense image features complicates spatiotemporal alignment. To address this, we propose DualRCS-BEV, a novel framework featuring three key innovations: (1) Dynamic RCS Prior Modeling: Leveraging Radar Cross-Section (RCS) values to generate geometrically constrained heatmaps encoding target size priors, with a learnable module adaptively predicting scattering ranges, effectively mitigating radar sparsity. (2) Dual-Branch RCS Prior Injection: A novel mechanism bidirectionally injects RCS geometric priors—enhancing radar BEV features and modulating image BEV features—significantly improving perception of target geometry and physical properties with minimal overhead. (3) Hierarchical Multi-Scale Fusion: A feature fusion strategy achieves precise radar-vision alignment and optimally balances near-field detail preservation with distant contour recognition. Extensive experiments on the nuScenes dataset validate the effectiveness of our innovations, demonstrating state-of-the-art performance in radar-camera 3D object detection.