SGFusion:稀疏感知门控的多模态三维目标检测OA

In autonomous driving perception,3D object detection often suffers from limited sensing modalities,leading to insufficient spa-tial modeling capability and degraded robustness,especially at long ranges and under heavy occlusions.Conventional single-modal meth-ods based on LiDAR or cameras alone struggle to provide stable and accurate detection in such challenging conditions.To address these issues,we propose SGFusion,a structure-aware multi-modal 3D object detection framework operating in Bird's Eye View(BEV)space.SGFusion integrates complementary LiDAR and image features to improve both accuracy and robustness.Specifically,a depth supervision signal derived from LiDAR point clouds is first introduced to provide explicit geometric priors for image BEV features,thereby enhancing the spatial consistency of cross-modal representations.On this basis,a sparsity-aware gated dual-attention mechanism(Sparsity-aware Gated Dual Attention,SGDA)is proposed to explicitly align geometric and semantic information between modalities in BEV space,while an input-dependent gating branch imposes sparsity on the cross-modal attention responses and adaptively reweights them according to the scene context.This design effectively suppresses depth-estimation artifacts and misalignment noise,and enhances robustness in complex driving scenes.Experiments on the nuScenes benchmark demonstrate that SGFusion achieves a mean Average Precision(mAP)of 74.9%and a nuScenes Detection Score(NDS)of 75.5%on the test set,outperforming representative single-modal and multi-modal baselines.