From Dense BEV to Sparse Queries: Efficiency and Design Trends in LiDAR-Camera Fusion for 3D Object Detection

Multi-modal LiDAR-camera fusion has become the dominant paradigm for 3D object detection in autonomous driving. Over the past four years, fusion architectures have shifted from dense Bird's-Eye View (BEV) representations toward sparse, query-based designs, motivated largely by computational efficiency. This survey presents a unified taxonomy of five categories: point-level fusion, dense BEV fusion, query- and proposal-level fusion, sparse fusion, and foundation-model-assisted fusion. It compares 20+ methods across nuScenes, KITTI, and Waymo Open, annotating GPU hardware, test-time augmentation, and image resolution for each entry. The paper also covers robustness under sensor degradation, temporal fusion, calibration-free fusion, end-to-end driving, and edge deployment. The analysis indicates that sparse fusion methods now achieve accuracy competitive with dense BEV approaches at substantially lower latency, establishing a new Pareto frontier for deployment-oriented design.