基于视觉的双分支结构鸟瞰图下的目标检测

Visual-Based Dual-Branch Structure for Bird’s-Eye-View Object Detection

摘要: 在鸟瞰图下基于视觉的3D目标检测中，会出现因伪点云生成过程中细节信息丢失而导致小目标和远距离目标检测性能不足以及计算量参数量较大的问题。为此，本文提出了一种基于视觉的双分支结构的3D目标检测方案。首先，采用跨阶段局部网络进行特征提取，以增强模型对多层次特征的学习能力并提升计算效率；其次，引入辅助特征提取网络，充分利用多尺度特征信息，强化对小目标和远距离目标的特征表达；最后，通过高效的特征融合策略，综合处理伪点云特征与辅助分支输出的特征，实现二维细节与三维信息的有效融合。实验结果表明，在nuScenes数据集上，所提方法的平均精度较BEVDet提升约2%，检测得分提升6.5%。由此可见，本文方案在兼顾计算效率的前提下，有效提高了3D目标检测精度，为无人驾驶系统的环境感知提供了有力支持。

Abstract: In vision-based 3D object detection under the bird’s-eye view, challenges arise due to the loss of fine details during the pseudo point cloud generation process, which results in suboptimal detection performance for small and distant objects, as well as high computational complexity and a large number of parameters. To address these issues, this paper proposes a vision-based dual-branch structure 3D object detection scheme. Firstly, a cross-stage partial network is employed for feature extraction to enhance the model’s ability to learn multi-level features while improving computational efficiency. Secondly, an auxiliary feature extraction network is introduced to fully leverage multi-scale feature information, thereby strengthening the representation of small and distant objects. Finally, an efficient feature fusion strategy is implemented to comprehensively integrate the pseudo point cloud features with the outputs from the auxiliary branch, effectively merging two-dimensional details with three-dimensional spatial information. Experimental results on the nuScenes dataset demonstrate that the proposed method achieves an approximately 2% improvement in mean average precision and a 6.5% increase in overall detection score compared to the BEVDet baseline. These findings indicate that the proposed framework not only enhances 3D object detection accuracy but also maintains computational efficiency, thereby providing robust support for the environmental perception capabilities of autonomous driving systems.