基于多视角二维投影的实景三维模型解译  

Classification of Textured 3D Mesh Models Based on Multiview 2D Mapping

作　　者：曲俸磊 胡忠文[1,2,3] 张英慧 张金华 邬国锋 QU Fenglei;HU Zhongwen;ZHANG Yinghui;ZHANG Jinhua;WU Guofeng(MNR Key Laboratory for Geo-Environmental Monitoring of Great Bay Area,Shenzhen University,Shenzhen 518060,China;Guangdong Key Laboratory of Urban Informatics,Shenzhen University,Shenzhen 518060,China;School of Architecture and Urban Planning,Shenzhen University,Shenzhen 518060,China)

机构地区：[1]深圳大学自然资源部大湾区地理环境监测重点实验室,深圳518060 [2]深圳大学广东省城市空间信息工程重点实验室,深圳518060 [3]深圳大学建筑与城市规划学院,深圳518060

出　　处：《地球信息科学学报》2024年第3期654-665,共12页Journal of Geo-information Science

基　　金：国家自然科学基金项目(41871227);深圳市科技计划资助(JCYJ20220818101617037、JCYJ20230808105201004)。

摘　　要：实景三维是对人类生产、生活和生态空间进行真实、立体化反映和表达的数字虚拟空间,已作为基础数据广泛应用于智慧城市、可视化展览等领域。实景三维模型语义解译是充分发掘实景三维模型的数据价值以实现场景的自动理解与分析的基础。已有的三维解译方法中,存在被遮挡地物解译不完全,不同地物边界解译不准确等问题。针对该问题,本研究提出了一种基于多视角的实景三维模型解译方法。该方法首先模拟成像过程实现实景三维的多视角二维投影,进一步基于二维影像解译快速获取语义信息,最后将多个二维解译结果进行逆投影获得三维语义模型。本研究以深圳大学实景三维模型为实验数据,通过与基于正射影像的三维解译、面向对象的三维分层解译等方法对比,验证本文所提方法的有效性。结果显示本研究提出的基于多视角的解译方法获得了最高的分类精度(总体分类精度为96.69%,Kappa系数为0.942),在正射遮挡区域,以及不同地物边界区域有更好的解译效果。本方法充分利用实景三维数据的多角度信息,具有较高的理论和实践参考价值,为进一步促进实景三维建设及在自然资源监测领域的应用提供了方法支撑。Textured 3D mesh models are digital virtual spaces that provide a true,three-dimensional representation of human production,living,and ecological spaces.They have been widely used as foundational data input in areas such as smart cities and visual exhibitions.The semantic interpretation of textured 3D models is the foundation for fully exploring the potential of these models to achieve automatic understanding and analysis of scenes.Existing interpretation methods suffer from issues such as incomplete interpretation of occluded objects and inaccurate interpretation of different object boundaries.To address these challenges,in this study,we propose a multiview-based classification method for textured 3D mesh models.A textured 3D mesh model is first segmented into ground surfaces and 3D objects by Cloth Simulation Filtering(CSF)method.The ground surface is projected to a 2D orthophoto and classified using object-based image analysis methods.The textured 3D objects are transformed into five 2D images through orthographic and multiview oblique projections.These 2D images are then classified using object-based image analysis methods.Furthermore,these 2D semantic maps are inverse-projected to the 3D mesh model,and a multiview voting strategy is proposed for fusing sematic information from different views to obtain the sematic 3D objects.Finally,the semantic terrain surface and 3D objects are merged together to obtain the semantic 3D mesh model.A textured 3D mesh model of Shenzhen University is used to verify the effectiveness of the proposed method.Besides,the proposed method is compared with two state-of-the-art methods.The results show that the proposed method effectively addresses the problems in interpreting occluded objects and distinguishing edges between different objects.It outperforms the competing methods,particularly in the areas of orthographic occlusion and where different ground objects are connected or adhered,and achieves the highest classification accuracy(overall accuracy is 96.69%,Kappa coefficient is

关 键 词：实景三维模型 三维解译 分层解译 多视角投影 面向对象 语义信息 视角加权投票 倾斜投影 

分 类 号：P282.1[天文地球—地图制图学与地理信息工程] P208[天文地球—测绘科学与技术]