UncertainBEV: Uncertainty-aware BEV fusion for roadside 3D object detection

With the rapid development of autonomous driving technology and intelligent transportation systems, multimodal fusion-based Bird’s-Eye-View (BEV) perception has become a key technique for environmental understanding. However, existing methods suffer from feature misalignment caused by calibration errors between different sensors, ultimately limiting the effectiveness of multimodal fusion. In this paper, we propose a robust roadside BEV perception framework, named UncertainBEV. To address feature misalignment caused by projection inaccuracies between LiDAR and camera sensors, we introduce a novel module called UncertainFuser, which models the uncertainty of both camera and LiDAR features to dynamically adjust fusion weights, thereby mitigating feature misalignment. Additionally, we optimize the sparse voxel pooling module and design a multi-head attention mechanism to enhance the quality of BEV features from both modalities. Built upon the CUDA-V2XFusion and BEVFusion frameworks, our proposed UncertainBEV achieves state-of-the-art performance on the DAIR-V2X-I dataset, with 3D mean Average Precision (mAP) improvements of 2.88%, 7.73%, and 3.68% for vehicles, pedestrians, and cyclists, respectively. Our code has been open-sourced at UncertainBEV.