Computer Engineering and Applications ›› 2023, Vol. 59 ›› Issue (9): 237-244.DOI: 10.3778/j.issn.1002-8331.2202-0075

• Graphics and Image Processing • Previous Articles     Next Articles

3D Object Detection Method Combining on Graph Sampling and Graph Attention

LI Wenju, CHU Wanghui, CUI Liu, SU Pan, ZHANG Gan   

  1. College of Computer Science and Information Engineering, Shanghai Institute of Technology, Shanghai 201418, China
  • Online:2023-05-01 Published:2023-05-01



  1. 上海应用技术大学 计算机科学与信息工程学院,上海 201418

Abstract: For the task of 3D object detection in point clouds, there are objects have a small scale or appear in complex scenes, which makes them have a lower detection accuracy. Therefore, a 3D object detection method based on graph sampling and graph attention mechanism using the point clouds is proposed. Firstly, the method reduces the size of down-sampling voxels to maintain the point clouds density of small objects, and then introduces graph sampling technology to reduce the cost of constructing topological graphs in the point clouds for feature extraction. Finally, the embedded self-attention mechanism in graphs before and after the graph sampling is used to enhance the feature extraction ability of the network. Compared with the benchmark on the KITTI dataset, proposed method improves the detection accuracy of car in hard scenes by 1.96%, and improves the detection accuracy of pedestrian and cyclist in moderate scenes and hard scenes with 4.21% and 2.57% respectively. Besides, the training time of proposed method is reduced by 15%. These demonstrate superior performance in detection accuracy of small objects in point clouds and the sampling method can improve the training efficiency of the model.

Key words: point cloud, 3D object detection, graph neural network, graph sampling, graph attention mechanism

摘要: 在点云中进行三维目标检测时,小目标和复杂背景下目标的检测精度不足是突出的问题之一。针对该问题,提出了一种基于图采样和图注意力机制的3D点云目标检测方法。减小基准网络下采样体素大小以保持小目标的点云密度;引入图采样降低在点云中构造拓扑图的代价;通过对图采样前后的图分别嵌入自注意力机制,提高网络的特征提取能力。在KITTI数据集上与基准网络Point-GNN相比,对汽车目标在复杂场景上的检测精度提升了1.96%,对行人与骑行者目标在中等难度场景和复杂场景上的检测精度分别提升4.21%和2.57%;与Point-GNN相比,减少了15%的训练时间。实验结果表明,设计的方法对于3D点云中小目标和复杂背景下目标的检测更加有效,图采样方法还能够提升模型的训练效率。

关键词: 点云, 3D目标检测, 图神经网络, 图采样, 图注意力机制