摘要
为改变目前部分三维点云数据集存在分辨率低、有噪声干扰和只有部分三维数据的问题,受最近神经网络成功的启发,提出一种基于深度学习的点云匹配方法。通过对RGB-D数据集重建获取三维点云数据集,利用卷积神经网络进行深度学习训练,得出一种能解决这些问题的描述子,完成点云匹配,使点云的匹配更加准确。实验结果表明,使用该方法进行点云匹配时,与其它点云匹配方法相比,具有更好的计算速度、运行时间更短,对点云旋转平移具有更好的鲁棒性,点云匹配的准确度更高。
To change the current low resolution,noisy interference and only partial 3D data of some 3D point cloud datasets,inspired by the recent success of neural networks,a point cloud matching method based on deep learning was proposed.The 3D point cloud dataset was acquired from the RGB-D dataset reconstruction,the convolutional neural network was used for deep learning training,a descriptor that solved these problems was derived and point cloud matching was completed.This method makes the matching of point clouds more accurate.The experimental results show that when using the proposed method for point cloud matching,it has higher calculation speed,shorter running time,better robustness to point cloud rotation and translation,and the accuracy of the matching of point clouds is higher.
作者
梁振斌
熊风光
韩燮
陶谦
LIANG Zhen-bin;XIONG Feng-guang;HAN Xie;TAO Qian(College of Computer Science and Technology,North University of China,Taiyuan 030051,China)
出处
《计算机工程与设计》
北大核心
2020年第6期1689-1693,共5页
Computer Engineering and Design
基金
山西省重点研发计划基金项目(201803D121081)
国家自然科学基金项目(61672473)。
关键词
三维点云
点云匹配
卷积神经网络
深度学习
描述子
3D point cloud
point cloud matching
convolutional neural network
deep learning
descriptor
