Occlusion-resilient 3D object detection from point cloud measurements for autonomous driving

Introduction In autonomous driving supported by perception modules, 3D object detection serves as a core task for achieving environmental understanding and safe decision-making. However, the occlusion problem severely compromises the completeness and geometric continuity of point cloud data especially in urban traffic environments. This issue significantly reduces the detection performance of models in partially and fully occluded scenarios. Although most existed 3D object detection models achieve high accuracy in non-occluded scenes, their feature representations become less effective in sparse point clouds or occluded areas, causing inaccurate canters or missing targets. Methods To overcome these limitations, this paper proposes MOVA (Multi-scale Occlusion-resilient Voxel Attention), a 3D detection method that enhances multi-scale sparse voxel features through a sparse key-selection attention module and a cross-scale attention fusion method. To further strengthen the learning of occluded instances, a DEIM-based query-matching strategy is used during training. Results On the nuScenes validation set, MOVA reaches a mean Average Precision (mAP) of 63.36%, outperforming other advanced methods. On the occlusion subset, MOVA reaches 61.44% mAP, 36.62 percentage points higher than the CenterPoint baseline (24.82%). Discussion These results confirm the effectiveness of MOVA in handling complex and highly occluded scenes,showing that focusing attention on informative voxel regions and integrating cross-scale context substantially improve the detection of occluded objects while maintaining real-time efficiency.