Inhaltszusammenfassung

Being able to measure the spatial motion of arbitrary objects with high accuracy and low latency is vital for numerous higher level tasks in many fields of application. These include, but are not limited to: robotic perception, medical navigation and mixed reality systems. Such measurements are typically obtained by consecutively estimating the object‘s pose, i.e. its location and orientation in three-dimensional space, relative to a known frame of reference. The most successful are approache...Being able to measure the spatial motion of arbitrary objects with high accuracy and low latency is vital for numerous higher level tasks in many fields of application. These include, but are not limited to: robotic perception, medical navigation and mixed reality systems. Such measurements are typically obtained by consecutively estimating the object‘s pose, i.e. its location and orientation in three-dimensional space, relative to a known frame of reference. The most successful are approaches based on optical sensors, such as digital cameras. But despite the large amount of literature and actively conducted research on this issue, fast, robust and accurate 3D object pose estimation still remains a key challenge in computer vision. This dissertation presents novel approaches to visual 3D object pose estimation from 2D images. The particular feature of the proposed solutions is that they operate in real-time while only requiring a single (monocular) camera. The main parts of this work describe an innovative active infrared LED marker-based system as well as a novel algorithm for passive markerless pose estimation, both developed within the course of this thesis. For the marker-based approach, two original, nearly co-planar LED patterns are proposed. These enable high-speed, single-image pose estimation of multiple markers as well as robustly avoiding common pose ambiguities. The proposed markerless method presents a novel combination of region-based and direct photometric pose estimation. It is enabled by a new numerical pose optimization strategy derived for the region-based part as well as an innovative statistical object segmentation model. The overall approach thereby significantly improved the robustness towards challenging conditions, such as dynamic lighting, cluttered backgrounds, different object appearances, occlusions and fast and complex motion, compared to the state of the art. It is furthermore the first capable of estimating the poses of multiple arbitrarily textured objects in real-time on a commodity laptop. In addition to this, a new complex dataset dedicated to the task of monocular object pose tracking has been created and made publicly available. Both proposed pose estimation solutions are extensively evaluated in numerous experiments, including the proposed as well as another popular public dataset. It is also shown that these solutions have been successfully applied in various practical scenarios, where they have enabled a variety of new problem solving opportunities.» weiterlesen» einklappen

Autoren

Klassifikation

DDC Sachgruppe:
Informatik