基于卷积神经网络的点云配准方法  

Point Cloud Registration Based on Convolutional Neural Network

作　　者：舒程珣 何云涛[1] 孙庆科 

机构地区：[1]北京航空航天大学电子信息工程学院,北京100191 [2]河北远东通信系统工程有限公司,河北石家庄050200

出　　处：《激光与光电子学进展》2017年第3期123-131,共9页Laser & Optoelectronics Progress

基　　金：上海航天创新基金SAST(2015090)

摘　　要：点云配准是三维点云信息处理中的重要问题。传统点云配准方法计算量大,不利于实时计算与移动计算。针对传统点云配准方法存在的问题,提出了一种利用卷积神经网络进行点云配准的方法。首先计算点云的深度图像,利用卷积神经网络提取深度图像对的特征差,将深度图像对的特征差作为全连接网络的输入并计算点云配准参数,迭代地执行上述操作直至配准误差小于可接受阈值。实验结果表明,相比传统的点云配准方法,基于卷积神经网络的点云配准方法具有所需计算量小、配准效率高、对噪声点和异常点不敏感的优点。Point cloud registration is an important issue in 3D information processing. The traditional point cloud registration needs a huge amount of computation, thus it is not suitable for real-time and mobile computation. In order to solve the problem of traditional point cloud registration method, a method based on convolutional neural network is proposed. The depth image of point cloud is calculated and the differential feature vector of depth images extracted by the convolutional neural network is regarded as input of fully connected neural network to calculate registration parameters. Iteratively executing the above process until registration error is acceptable. Experimental results show that the point cloud registration based on convolutional neural network is simpler in computation, maro efficient in registration rate, and less sensitive to noise and outlier than the traditional methods.

关 键 词：图像处理 点云配准 深度学习 卷积神经网络 深度图像 

分 类 号：TP249[自动化与计算机技术—检测技术与自动化装置]