Benchmarking Robustness-Efficiency Trade-offs of Camera-LiDAR Fusion Strategies for 3D Object Detection Under Environmental Corruptions

Multi-modal sensor fusion combining camera and LiDAR data has become the dominant paradigm for 3D object detection in autonomous driving. The selection among early fusion, late fusion, and deep fusion strategies involves complex trade-offs between detection accuracy, computational efficiency, and robustness under adverse conditions. This paper presents a systematic benchmarking study that quantitatively evaluates these trade-offs on the nuScenes dataset under diverse environmental corruptions and calibration perturbations. Six representative fusion algorithms spanning three fusion categories are evaluated across 10 corruption types at three severity levels, with simultaneous measurement of detection accuracy (mAP, NDS), computational resource consumption (latency, GPU memory), and degradation patterns under spatial misalignment and temporal desynchronization. The results reveal that deep fusion approaches achieve 2.8%--5.1% higher NDS than early fusion under clean conditions, while late fusion strategies demonstrate 12.3%--18.7% lower mean Corruption Error under LiDAR-degraded scenarios. Calibration perturbation analysis shows that soft-association mechanisms reduce mAP degradation by 41.2% compared to hard-association approaches at 0.5-meter spatial misalignment. These findings provide evidence-based guidance for engineering teams selecting sensor fusion configurations under real-world deployment constraints.