面向对象的实景三维模型解译研究

The rapid development of UAV remote sensing platform and 3D scene reconstruction technology have brought revolutionary progress on earth observation technology. The textured 3Dscene model is one of the most important products of UAV remote sensing, showing great potential in many applications. However, it is difficult to fully explore the advantages of 3D scene model, due to the underdeveloped interpretation techniques. Textured 3D models are usually converted to orthophotos and DSMs, and then used for feature extraction.. In this study, we are aiming at developing the theory and technique for object-based 3D scene model interpretation. In our framework, the textured 3D scene models will be segmented in to 3D objects and then be interpreted directly. It is not necessary to converted the model to traditional 2D data. It is expected that less information is lost, and it is more convenient and reliable to extract information from 3D model directly.. This research contains four aspects:. (1) The modeling of the hierarchical and multi-scale distribution of different landscapes. In this study, the distributions of landscapes over different scenes are studied and modelled. These information is used as reference for multi-scale 3D model segmentation and interpretation.. (2) Multiple feature embedded 3D model segmentation. In this study, the spectral, textural and shape features of the scene model are modelled and fused. And then, a region merging based segmentation algorithm is used to construct a region-based hierarchy, where the 3D objects are represented as a node in this hierarchy.. (3) The selection of optimal scale parameters for different landscapes. In this study, the optimal scale for each landscape is studied, combining the spectral, textural and shape features. The corresponding segmentation results are retrieved from the region-based hierarchy.. (4) The hierarchical interpretation of different landscapes. The hierarchical relationships are explored, and the different landscapes are interpreted hierarchically. Sever typical scenes will be used to evaluate the performance of the whole framework.. It is expected that the textured 3D scene model could be interpreted directly, without transfering to 2D image data. And furthermore, the efficiency, accuracy and automation of 3D scene model analysis can be significantly improved. This research therefore can help us to better take advatages of 3D scene models, and promote the usage in urban, algriculture, forest and ecology application.

随着无人机遥感平台和倾斜摄影测量技术的快速发展，实景三维模型这一新型遥感数据在诸多方面展现出巨大优势和应用潜力，然而三维模型解译技术发展相对滞后，使之难以发挥出潜在的价值。本课题拟从信息提取的角度，以面向对象的影像分析技术为基础，研究实景三维模型中多层次、多尺度的模型分割与信息提取，实现“三维模型->三维信息”的直接解译。本项目的研究内容包括：1.首先研究三维环境下地物的多层次、多尺度分布规律，并对其进行定量表达；2.在此基础上开展三维模型的多层次、多尺度分割，将模型转化为基于区域的多层次结构；3.根据地物的分布规律，研究不同地物的最优分割尺度选择；4.根据地物提取的多层次、多尺度策略，逐步实现三维模型的地物提取，进行验证实验和反馈。本课题的研究成果对推动三维模型解译理论和技术的发展，推进在其在城市、农林业、生态、灾害等领域的应用具有积极的意义。

实景三维模型是当前无人机载倾斜摄影测量技术的主要测绘产品，在城市管理、资源环境调查等方面发挥越来越多的作用。然而实景三维模型的解译仍然处于起步阶段，相关技术和方法均不完善。为此，课题组按照原定计划开展了实景三维模型地物立体分布规律与分层解译策略、实景三维模型层次化、多尺度分割方法、实景三维模型多源特征提取与选择、实景三维模型语义分类等研究，在深圳大学校园及周边、广州科学城、深圳市福田内伶仃自然保护区等典型场景采集包括实景三维模型、多/高光谱影像、点云等多源数据，开展相应的应用研究。在项目的支持下，研究团队探索并验证了面向对象的实景三维模型分层解译策略和方法，探索了实景三维模型解译在城市环境监测、山地入侵植被检测、海岛礁植被精细分类等方面的应用潜力和优势。..项目研究团队发表了一系列高水平论文，其中SCI论文12 篇，中文论文1篇，相关研究成果在国际会议口头报告2次，国内会议5次，申请发明专利7项，形成了原型软件系统1套，申请了软件著作权1项，培养博士后2人，博士生2人，硕士生5人，获得测绘科技进步奖3项，地理信息科技进步奖1项，广东省自然科学二等奖1项，总体上达到了预期的目标。