Eva Stergaršek Kuzmič (2010) Detecting new objects and building models with active robot system. MSc thesis.
An important element of a cognitive robotic system is the ability to detect novel objects and learn their representations, which are suitable for later recognition and manipulation. The basic assumption of our work is that the detection and segmentation of new objects can be facilitated by an active robotic system, which can not only observe the objects but can also manipulate them. Manipulation supports object segmentation and the accumulation of object features, which provides the basis for building object models and for determining their functionality. In this thesis we propose a new approach for object detection and segmentation based on the integrated use of visual and manipulative functions of a robotic system. The developed system was tested in a number of experiments on a real robot. In the proposed approach, the object detection process consists of three consecutive procedures. The first procedure deals with the generation of hypotheses about the existence of an object. It comprises several sub-processes: extraction of visual features, calculation and clustering of 3-D points, and discovery of planar surfaces. The second process defines manipulative actions which the robot needs to perform to validate the previously calculated object hypotheses. By employing these manipulative actions, the existence of an object can be confirmed and additional object knowledge can be accumulated. In our system, suitable manipulative actions are realized as pushing movements. After the completion of the pushing movement, the third procedure verifies the underlying object hypothesis based on the newly acquired information. The verification procedure evaluates the consistency of the detected object features with respect to the assumption of rigid body motion. Additional hypotheses are evaluated if the verification process fails. To ensure the robustness of the system, probabilistic methods such as RANSAC (RANdom SAmple Consensus) have been applied in several computational stages. Our experimental results show that the system provides successful object detection and segmentation in complex scenes. Our work also demonstrates the robot can acquire more complete object models by performing several consecutive manipulative actions.
Actions (login required)