MF-BEVFusion: multiscale depth estimation and fully dynamic fusion for camera-LiDAR BEV 3D object detection

Object detection is a pivotal task in autonomous driving, where camera-LiDAR fusion is widely adopted to balance semantic richness and geometric accuracy. However, existing methods are often impeded by feature misalignment in Bird’s-Eye-View (BEV) space, originating from the inherent depth limitations of cameras and the semantic sparsity of LiDAR. In addition, the neglect of dynamic modal interaction further exacerbates information loss. To overcome these challenges, we propose MF-BEVFusion, a novel multimodal fusion framework. Unlike methods that rely exclusively on camera-based depth prediction, we introduce a LiDAR-guided multiscale depth estimation (MDE) module. This module explicitly incorporates sparse LiDAR geometric cues to rectify image depth distribution, effectively resolving semantic misalignment in BEV space. Furthermore, addressing the limitations of static concatenation, we design a fully dynamic fusion (FDF) module. By leveraging global context, the FDF module adaptively recalibrates feature importance, ensuring robust detection even under occlusion or poor lighting. Experimental results on the nuScenes dataset demonstrate that our method outperforms baseline algorithms, achieving improvements of 1.12% in mean average precision (mAP) and 1.03% in NuScenes detection score (NDS). Moreover, generalization experiments on the KITTI dataset verify the algorithm’s stability and effectiveness in complex autonomous driving environments.