摘要:
随着影像密集匹配方法的发展,目前可以从多视倾斜航空影像获得大量类比于激光扫描数据密度甚至精度的点云,但获取结果以着色的点云为主,缺乏分类信息.针对此问题,提出了一种面向对象的倾斜摄影测量点云分类方法.首先,计算单点特征向量;然后,利用SLIC(simple linear iterative clustering)算法将点云对应的影像分割成超像素,再根据点云和影像间的关系,将点云聚类成超体素对象,并计算每个对象的特征向量;在此基础上,采用随机森林算法对超体素进行分类;最后,根据语义信息对分类结果进行后处理获得最终的点云分类结果.2组典型实验数据结果表明,总体分类精度分别达到91.2%和88.1%,比基于单点的分类方法分别提高了2.3%和8.2%.
Abstract:
With the development of image dense matching method,point clouds can be obtained from multi-view oblique aerial images, whose accuracy and density can be comparable with LiDAR point clouds. However, the currently derived colored point clouds lack classification information. In view of such a situation, this paper proposes an object -based classification method for oblique photogrammetric point clouds. The first step of this method is to calculate features of each point. Then,SLIC algorithm is used to divide the corresponding image into super-pixels. After that, point clouds are clustered into super -voxels as objects according to the relationship between point clouds and images,and features of each object are calculated afterwards. Random forests algorithm is used to classify these super-voxels. Finally, contextual information is adopted to optimize the initial classification results. Two sets of data were employed for evaluating the proposed method, and the overall accuracy could reach up to 91.2% and 88.1% respectively,which improves the precision by 2.3% and 8.2% compared with the point-based classification.
下载: