摘要
目前的三维目标检测方法大多采用最远点采样进行特征提取,但忽略了点云密度信息所带来的作用。为了进一步增强稀疏卷积能力与提升目标特征信息量,提出一种动态稀疏卷积与点密度感知融合的网络(DS-PDP),首先对点云体素化经过动态稀疏卷积下采样,通过体素点质心定位从动态稀疏卷中定位体素特征信息,其次利用核密度估计方法,使感兴趣区域网格池对点质心密度特征多尺度聚合,最后对目标置信度预测,并在KITTI数据集上取得了有竞争力的结果。
Most of the current 3D object detection methods use the farthest point sampling for feature extraction,but ignore the effect of point cloud density information.In order to further cnhancc the ability of sparse convolution and improve the amount of targct featurc information,proposes a dynamic sparse convolution point density perception network(DS-PDP).Firstly,the point cloud is voxelized by dynamic sparse convolution down-sampling,and the voxel feature information is located from the dynamic sparse volume by voxel point centroid localization.The grid pool of the region of interest is used to multi-scale aggregate the point centroid density fea-tures,and finally the target confidence is predicted,and competitive results are achieved on the KITTI dataset.
作者
张秀清
赵泽洋
许云峰
ZHANG Xiuqing;ZHAO Zeyang;XU Yunfeng(School of Infomation and Engineering,Hebei University of Science and Technology,Shijiazhuang 050018,China)
出处
《长江信息通信》
2024年第5期61-64,共4页
Changjiang Information & Communications
基金
河北省重点研发计划项目资助项目(No.21373802D)
教育部人工智能协同育人项目(No.201201003011)。
关键词
动态稀疏卷积
体素点质心
点密度
ROI
Dynamic sparse convolution
Voxel point centroid
Point density
Rol
