【背景】传统方法因静态感受野设计较难适配城市自动驾驶场景中汽车、行人及骑行者等目标的显著尺度差异，且跨尺度特征融合易引发层级干扰。【方法】针对自动驾驶场景中多类别、多尺寸目标的3D检测中跨尺度表征一致性的关键挑战，本研究提出基于均衡化感受野的3D目标检测方法VoxTNT，通过局部-全局协同注意力机制提升检测性能。在局部层面，设计了PointSetFormer模块，引入诱导集注意力模块（Induced Set Attention Block, ISAB），通过约简的交叉注意力聚合高密度点云的细粒度几何特征，突破传统体素均值池化的信息损失瓶颈；在全局层面，设计了VoxelFormerFFN模块，将非空体素抽象为超点集并实施跨体素ISAB交互，建立长程上下文依赖关系，并将全局特征学习计算负载从O（N ²）压缩至O（M ²）（M<<N， M为非空体素数量），规避了复杂的Transformer 直接使用在原始点云造成的高计算复杂度。该双域耦合架构实现了局部细粒度感知与全局语义关联的动态平衡，有效缓解固定感受野和多尺度融合导致的特征建模偏差。【结果】实验表明，该方法在KITTI数据集单阶段检测下，中等难度级别的行人检测精度AP（Average Precision）值达到59.56%，较SECOND基线提高约12.4%，两阶段检测下以66.54%的综合指标mAP（mean Average Precision）领先次优方法BSAODet的66.10%。同时，在WOD数据集中验证了方法的有效性，综合指标mAP达到66.09%分别超越SECOND和PointPillars基线7.7%和8.5%。消融实验进一步表明，均衡化局部和全局感受野的3D特征学习机制能显著提升小目标检测精度（如在KITTI数据集中全组件消融的情况下，中等难度级别的行人和骑行者检测精度分别下降10.8%和10.0%），同时保持大目标检测的稳定性。【结论】本研究为解决自动驾驶多尺度目标检测难题提供了新思路，未来将优化模型结构以进一步提升效能。

[Background] Traditional methods, due to their static receptive field design, struggle to adapt to the significant scale differences among cars, pedestrians, and cyclists in urban autonomous driving scenarios. Moreover, cross-scale feature fusion often leads to hierarchical interference. [Methodology] To address the key challenge of cross-scale representation consistency in 3D object detection for multi-class, multi-scale objects in autonomous driving scenarios, this study proposes a novel method named VoxTNT. VoxTNT leverages an equalized receptive field and a local-global collaborative attention mechanism to enhance detection performance. At the local level, a PointSetFormer module is introduced, incorporating an Induced Set Attention Block (ISAB) to aggregate fine-grained geometric features from high-density point clouds through reduced cross-attention. This design overcomes the information loss typically associated with traditional voxel mean pooling. At the global level, a VoxelFormerFFN module is designed, which abstracts non-empty voxels into a super-point set and applies cross-voxel ISAB interactions to capture long-range contextual dependencies. This approach reduces the computational complexity of global feature learning from O(N²) to O(M²) (where M << N, M is the number of non-empty voxels), avoiding the high computational complexity associated with directly applying complex Transformers to raw point clouds. This dual-domain coupled architecture achieves a dynamic balance between local fine-grained perception and global semantic association, effectively mitigating modeling bias caused by fixed receptive fields and multi-scale fusion. [Results] Experiments demonstrate that the proposed method achieves a single-stage detection Average Precision (AP) of 59.56% for moderate-level pedestrian detection on the KITTI dataset, an improvement of approximately 12.4% over the SECOND baseline. For two-stage detection, it achieves a mean Average Precision (mAP) of 66.54%, outperforming the second-best method, BSAODet, which achieves 66.10%. Validation on the WOD dataset further confirms the method’s effectiveness, achieving 66.09% mAP, which outperforms the SECOND and PointPillars baselines by 7.7% and 8.5%, respectively. Ablation studies demonstrate that the proposed equalized local-global receptive field mechanism significantly improves detection accuracy for small objects. For example, on the KITTI dataset, full component ablation resulted in a 10.8% and 10.0% drop in AP for moderate-level pedestrian and cyclist detection, respectively, while maintaining stable performance for large-object detection. [Conclusions] This study presents a novel approach to tackling the challenges of multi-scale object detection in autonomous driving scenarios. Future work will focus on optimizing the model architecture to further enhance efficiency.

利益冲突：Conflicts of Interest 所有作者声明不存在利益冲突。

All authors disclose no relevant conflicts of interest.