Intuitively Trained Segmentation of Industrial Workpieces for Robotic Manipulation

Kurzfassung

The research and application field robotics consists of a wide range of complex subtasks that all need to work together as a unit to produce a secure, reliable and efficient machine humans can work with. This thesis presents a solution to robotic vision using an intuitively trained segmentation as basis for robotic manipulation of industrial workpieces. Features, such as sharp edges and corners, are common keypoints to look for if the objects provide surface textures. However, industrial workpieces lack of those textures and are often metallic what causes the reflection of light to have a big impact on the objects appearance. Hence, the method described in this thesis is not based on edges but uses geometrical primitives to detect an object. The relations among those primitives play a key role in detecting the position of an object. They build a so called similarity matrix which represents the matching behaviour of primitive pairs. This matrix is then used in a subsequent segmentation process to assign detected primitives to corresponding objects in the image. Although most detection approaches depend on a high number of training data to achieve good results, the method proposed in this thesis gets a long with a few images. The results in chapter 5 show that two arbitrarily positioned objects can be clustered correctly with only 9 training images.