面向对象的城市道路及周边车载激光扫描点云分类理论和方法

Mobile Laser Scanning (MLS) is one of the most advanced techiques in the field of suveying and mapping,and it has wide applications such as large scale relief map update,infrastructure survey, 3D reconstruction of street view and 3D navigation etc. MLS point cloud classification is a basic and nessary processing step for the varios types of applications. However, the exisint classification methods fail to make use of the coplanar features of the discrete points and priori knowledge about the various types of objects. Herein, an object-based classifiction thoery is proposed to classify the MLS point clouds based on the idea of object-based image analysis in the field of remote sensing. In our proposed thoery, priori knowledge is tranfered into some features and rules; point cloud segmentation is employed to separate the discrete points into objects with sematics; the vehile trajectory,geometric, shape, radiometric and topological features are calculated, and they are used to classify the segmented objects into the ground, fa?ade, building roof,fence, road guardrail, traffic light, traffic sign, vehilce and human beings with a fixed order; at last ,the classification results are opitimized based on the prioir knowlege. In the above reserch, the basis thoeries about the optimal and adaptive parameters selection meodel for segmentation, expression of prioir knowledge, 3D outlines extraction of regular and irregular shapes of objects are focused on, and the new methods about object-based filtering, object-based fa?ade extraction and object-based fuzzy classification of clusted points are proposed, which is aimed to form a new classification thoery for MLS point clouds.

车载激光扫描是当今测绘界最前沿的科技之一，在大比例尺地形图修测、城市部件普查、三维街景重建和导航等领域具有重要价值。其中，点云分类是数据处理的重要组成部分。已有点云分类方法存在没有充分利用点云数据本身反应的面状几何特征和地学先验知识等弊端。本申请充分借鉴面向对象的遥感影像分析的理论和方法，以先验知识为引导，以点云分割为基础，充分利用车辆行驶轨迹线、点云数据的几何、形状（形态）、辐射、拓扑关系等特征，按照一定的顺序进行地面点、建筑物立面点、屋顶点、道路护栏网、围墙（篱笆）、道路交通灯、交通标志牌、车辆、行人等目标的识别，并利用先验知识进行目标识别结果的修正和优化。研究过程中，侧重自适应分割参数的选取模型、先验知识的特征表达方式、地物规则和不规则三维轮廓线提取等基础理论和方法研究，围绕面向对象的点云滤波、建筑物立面提取、模糊分类等创新点，建立一套面向对象的车载激光扫描点云分类理论和方法。

点云的自动化处理是高精度地图生产的必经环节。但，迄今为止，自动化、智能化点云处理和信息提取，尤其是点云分类和目标识别，又是高精度地图生产的难点和瓶颈。本项目主要围绕点云多尺度分割、滤波、道路交通标记线提取、路肩提取、杆状物提取、隔离路障提取、广告牌提取、交通信号灯横杆提取、天桥提取、建筑物立面提取、交通指示牌提取、车辆提取、垃圾桶提取、电力导线提取、树叶识别、林木参数半自动提取、粗差检测等做了深入研究，研发了相应的算法，并开发了点云处理和信息提取的的原型系统点云催化剂（Point Cloud Catalyst，PCC）。PCC软件包含的17个功能模块，每个模块不需要、或至多需要1-2个参数，且通常情况下使用参数缺省值也能取得很好的点云处理效果，克服了目前既有软件和相关算法需要大量人工干预和参数设置的弊端。目前，装配了该软件的单台计算机工作站可以在24小时内完成500亿个点的处理和信息提取任务，且在无需训练样本的情况下实现点云自动分类和目标识别精度均在90%以上，PCC软件具有高自动化、高智能、高效、高精度的优点。另外，PCC软件还具有支持批量处理、处理的单个点云文件的大小不受算法的限制、仅受内存大小的限制的特点；另外，该软件分类和目标识别结果均保存为las格式的点云标准文件、矢量化的结果保存为shp、或者dwg格式的矢量格式文件，便于与任何其他第三方点云处理、可视化和编辑软件的无缝衔接。另外，项目发表论文27篇，其中SCI检索8篇、EI检索8篇、中文核心11篇；项目获得发明专利3项；申请软件著作权1项；出版学术专著1部。圆满完成了既定的各项考核指标。