Frustumpillar-Fusion: Gated 4D Radar-Camera Fusion for USVs

Robust object detection for Unmanned Surface Vehicles (USVs) in inland waterways remains challenging due to projection-induced depth collapse and Doppler cancellation in radar-camera fusion. We propose FrustumPillar-Fusion, an end-to-end 4D radar-camera detector that removes explicit radar-to-image projection. A Frustum Pillar Encoder performs ray-aligned 3D partition to preserve geometric and Doppler cues, while a Radar-Gated Fusion Module uses a pixel-wise map α and a global scalar β to decouple spatial blending from residual refinement. On WaterScenes, our method achieves 65.0 mAP_50-95 and 91.3 mAP₅₀, outperforming listed baselines under our current benchmark setting and showing stronger robustness in challenging scenes. These results indicate that preserving frustum-aligned radar structure before cross-modal interaction is essential for reducing early feature contamination. The proposed design further provides a practical radar-camera fusion strategy for safety-critical USV perception in complex inland waterways.