动态卷积的3D点云目标检测算法  

3D Point Cloud Target Detection Algorithm Based on Dynamic Convolution

作　　者：蔡黎明 杜吉祥[1,2,3] 刘怀进 张洪博[1,2,3] 黄敬东[1,2,3] CAI Liming;DU Jixiang;LIU Huaijin;ZHANG Hongbo;HUANG Jingdong(College of Computer Science and Technology,Huaqiao University,Xiamen 361021,China;Fujian Key Laboratory of Big Data Intelligence and Security,Huaqiao University,Xiamen 361021,China;Xiamen Key Laborstory of Computer Vision and Pattern Recognition,Huaqiao University,Xiamen 361021,China)

机构地区：[1]华侨大学计算机科学与技术学院,福建厦门361021 [2]华侨大学福建省大数据智能与安全重点实验室,福建厦门361021 [3]华侨大学厦门市计算机视觉与模式识别重点实验室,福建厦门361021

出　　处：《华侨大学学报（自然科学版）》2023年第1期111-118,共8页Journal of Huaqiao University(Natural Science)

基　　金：国家自然科学基金资助项目(61673186);国家重点研发计划专项项目子课题(2019YFC1604705);福建省厦门市科技计划项目(3502220193037)。

摘　　要：针对不规则且稀疏的点的提取特征问题,提出一种以动态卷积作为特征提取的3D点云目标检测算法.首先,以一种新型的动态卷积的方式自适应学习点的位置特征,分类出前景点与背景点,同时对提取出的前景点逐一做回归框;然后,用非极大值抑制选出分数值最好的回归框.其次,进行粒度的细化,得到修正规范的3D回归框,完成3D物体的目标检测.最后,在KITTI数据集上验证算法的有效性.结果表明:文中所提算法在汽车类、行人类、自行车类数据集上的3D点云目标检测精度更高.Aiming at the problem of extracting features from irregular and sparse points,a 3D point cloud target detection algorithm using dynamic convolution as feature extraction is proposed.Firstly,a new dynamic convolution method is used to adaptively learn the position features of points,and classify the foreground points and background points.At the same time,the extracted foreground points are used as regression boxes one by one.Then,non maximum suppression is used to select the regression box with the best score,secondly,the granularity is refined to obtain a 3D regression box of the revised standard,and the target detection of 3D objects is completed.Finally,the validity of the algorithm is verified on KITTI data set.The results show that the proposed algorithm has higher detection precision of 3D point cloud target on the car,pedestrian and bicycle data sets.

关 键 词：点云 3D目标检测 动态卷积 分类回归 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]