融合体素图注意力的三维目标检测算法  

3D object detection algorithm with voxel graph attention

作　　者：鲁斌[1,2] 孙洋 杨振宇 LU Bin;SUN Yang;YANG Zhenyu(School of Control and Computer Engineering,North China Electric Power University,Baoding 071003,China;Engineering Research Center of Intelligent Computing for Complex Energy Systems,Ministry of Education,Baoding 071003,China)

机构地区：[1]华北电力大学控制与计算机工程学院,河北保定071003 [2]复杂能源系统智能计算教育部工程研究中心,河北保定071003

出　　处：《智能系统学报》2024年第3期598-609,共12页CAAI Transactions on Intelligent Systems

基　　金：国家自然科学基金项目(62371188);河北省在读研究生创新能力培养项目(CXZZBS2023153).

摘　　要：目前基于点云的三维目标检测方法未能充分利用点云局部几何特征,导致对点云稀疏的目标检测效果不佳。为此,本文提出基于原始点云体素图注意力的两阶段三维目标检测算法(voxel graph attention region-CNN,VGT-RCNN)。通过多尺度体素特征插值计算网格中心点特征;在多尺度非空体素特征上构造局部图;通过图注意力机制对体素特征进行加权平均,充分提取并利用目标的局部几何特征完成检测。该算法主要针对当前二阶段算法在进行特征聚合时对不同体素特征的重要性考虑不足进行改进,引入可学习的权重矩阵,动态学习体素特性的权重,提高局部特征表达能力。本文在流行的KITTI自动驾驶数据集上进行了充分测试,取得了具有竞争力的检测效果,尤其是在对点云稀疏的汽车目标检测上,准确率有较大提高。本文还对检测效果进行了可视化分析。Current point cloud-based 3D object detection methods fail to fully use the local geometric features of the point clouds,leading to poor performance in detecting objects of sparse point clouds.To solve this problem,a two-stage 3D object detection algorithm named voxel graph attention region-CNN(VGT-RCNN)is proposed based on the voxel graph attention of raw point clouds.Initially,the grid center point features are calculated by multiscale voxel feature interpolation.Then,a local graph is constructed on the multiscale non-empty voxel features.Finally,a weighted average is conducted for the voxel features by graph attention mechanism,fully extracting and using the local geometric features of the object to complete detection.The algorithm mainly improves the defect of the present two-stage algorithm,which does not sufficiently consider the significance of different voxel features in feature clustering.In addition,a learnable weight matrix is introduced to dynamically learn the weight of the voxel feature and increase the expression ability of local features.The algorithm has been sufficiently tested on the popular KITTI autonomous driving dataset,obtaining competitive detection effects.The accuracy of cars with sparse point clouds has been markedly improved.A visualized analysis is also carried out to determine the detection effect.

关 键 词：点云 三维目标检测 图注意力 特征插值 多尺度特征 激光雷达 体素化 车辆检测 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]