Robust machine vision for service robotics

In this thesis the vision architecture ROVIS (RObust machine VIsion for Service Robotics) is suggested. The purpose of the architecture is to improve the robustness and accuracy of visual perceptual capabilities of service robotic systems. In comparison to traditional industrial robot vision where the working environment is predefined, service robots have to cope with variable illumination conditions and cluttered scenes. The key concept for robustness in this thesis is the inclusion of feedback structures within the image processing operations and between the components of ROVIS. Using this approach a consistent processing of visual data is achieved. Specific for the suggested vision system are the novel methods used in two important areas of ROVIS: definition of an image ROI, on which further image processing algorithms are to be applied, and robust object recognition for reliable 3D object reconstruction. The ROI definition process, build around the well known „bottom-up top-down“ framework, uses either pixel level information to construct a ROI bounding the object of interest or contextual knowledge from the working scene for bounding certain areas in the imaged environment. The object recognition and 3D reconstruction chain is developed for two cases: region and boundary based detected objects. Since vision in ROVIS relies on image segmentation on each processing stage, that is ROI definition and object recognition, robust segmentation methods had to be developed. As said before, the robustness of the proposed algorithms, and consequently of ROVIS, is represented by the inclusion of feedback mechanisms at image processing levels. The validation of ROVIS is performed through its integration in the overall control architecture of the service robotic system FRIEND. The performance of the proposed closed-loop vision methods is evaluated against their open-loop counterparts.