基于语义分割的城区激光点云与光学影像配准  

Registration of Laser Point Cloud and Optical Image in Urban Area Based on Semantic Segmentation

作　　者：朱颖 赵明[1,2] ZHU Ying;ZHAO Ming(College of Information Engineering,Shanghai Maritime University,Shanghai 201306,China;Key Laboratory of Intelligent Infrared Perception,Chinese Academy of Sciences,Shanghai 200083,China)

机构地区：[1]上海海事大学信息工程学院,上海201306 [2]中国科学院智能红外感知重点实验室,上海200083

出　　处：《光子学报》2021年第1期188-202,共15页Acta Photonica Sinica

基　　金：上海市自然科学基金面上项目(No.20ZR1423500);中国科学院智能红外感知重点实验室开放课题(No.CAS‒IIRP‒04);中科院先导项目(No.XDA13030402)。

摘　　要：基于点云数据与光学遥感影像的协同应用在遥感领域获得广泛关注,为了对两种数据进行精确的配准以更好地融合两者优势,提出了一种城市场景下点云与光学遥感影像的自动配准方法。首先,由点云数据生成深度影像,即3D数据转换为2D影像;然后,运用Unet模型对深度影像和光学遥感影像分别进行训练并分割得到建筑面;再基于建筑面轮廓点集构建建筑最小外接矩形,将矩形长宽比作为寻找同名点的约束条件;接着,利用相似三角形原理寻找矩形中心同名点;最后,同名点坐标代入变换模型计算模型参数,完成配准。实验结果表明该方法对于运用传统点特征方法匹配困难的情况可实现较好的配准效果,且对图像平移、旋转、缩放均具有可抗性。The collaborative application of point cloud data and optical remote sensing image has been widely concerned in the field of remote sensing.In order to accurately register two kind of data and better integrate their advantages,an automatic registration method of point cloud and optical remote sensing image in urban scene is proposed.Firstly,the depth image is generated from point cloud data,that is,3D data is converted into 2D image.Secondly,the Unet model is used to train the depth image and the optical remote sensing image respectively and get building segmentations.Thirdly,the minimum circumscribed rectangles of buildings are constructed based on the contour set of building segmentation,and the lengthwidth ratio of rectangle is taken as the constraint condition to find Corresponding Points(CPs).Then,we use the similar triangle principle to find CPs of the rectangle’s center point.Finally,the coordinate of the CPs are substituted into the transformation model to calculate the model parameters,thus the registration is achieved.The experimental results show that the proposed method can achieve better registration effect when it is difficult to match with traditional point feature method,and it is resistant to image translation,rotation and scaling.

关 键 词：遥感 配准 语义分割 点云 光学 

分 类 号：TP751[自动化与计算机技术—检测技术与自动化装置]