Online CAD reconstruction with hand-held depth cameras (HandCAD-2)

三维几何计算机模型（CAD模型）用于许多技术应用，如建筑、制造、仿真和计算机游戏。这些模型服务于不同的目的，但有一个共同点，即它们代表了对现实世界中刚体的模仿。CAD模型是在设计师的几个复杂步骤中手动创建的，或者是在计算机上的一个创造性过程中创建的。或者，由传感器捕获现有物体或现有环境，然后从测量（CAD重建）中自动生成几何模型。今天，低成本的深度相机可用于CAD重建。除了这样收集的几何数据之外，还需要检测对象的附加抽象信息，例如其属性、对称性或特征。此外，即使没有专业知识，CAD重建也应该快速且易于执行。在本研究项目中，开发和研究了从手持相机深度传感器数据自动生成CAD模型的方法。在这里，考虑以不同分辨率级别的抽象模型形式对单个对象和完整的内部或外部场景进行在线可计算重建。应该使用低性能的计算硬件生成重建，以便在获取过程中已经显示结果。具体的挑战是非平面的、无结构的或对称的表面和动态物体，比如在场景中移动的人。从长远来看，对重建的物体得出额外的结论是可取的，例如：它是哪个物体？物体是固定的还是可移动的？物体（例如一扇门）可能的运动是什么？

Three-dimensional geometric computer models (CAD models) are used in many technical applications, such as in construction, manufacturing, simulation and in computer games. These models serve very different purposes, but have in common that they represent an imitation of rigid bodies in the real world. A CAD model is created either manually in several complex steps of a designer, or in a creative process on the computer. Alternatively, an existing object or an existing environment is captured by sensors and then a geometric model is automatically generated from the measurements (CAD reconstruction). Today, low-cost depth cameras are available for CAD reconstruction. In addition to the geometric data thus collected, additional abstract information of the detected objects, such as their properties, symmetries or features are desirable. Furthermore, the CAD reconstruction should be quick and easy to perform even without expert knowledge.In this research project methods for the automatic generation of CAD models from sensor data from hand-held cameras depth are developed and investigated. Here, the online computable reconstruction of individual objects and of complete interior or exterior scenes in the form of abstract models is considered with different resolution levels. The reconstruction should be generated with low performance computing hardware, such that the results can be displayed already during the acquisition. Specific challenges are non-planar, structure-less or symmetric surfaces and dynamic objects, such as people moving through the scene. In the long run, it would be desirable to draw additional conclusions on the reconstructed objects, such as: Which object is it? Is the object fixed or movable? What are the possible movements of the object (e.g. a door)?