In the fully automated assembly line, picking up the parts out of bin is one of the essential tasks for the part feeding job. Robot vision is the most powerful sensing tool for this bin picking task. In this thesis, we present the vision algorithms to address three major issues of the bin picking task: low-level feature extraction, part isolation, and grasp planning. An idea of using the frame coherence between the successive pick operations is also presented. We restrict the parts to the ones having the cylindrical shape.
For the purpose of edge enhancement in the low-level feature extraction, we present two new techniques: multiple light sources and gap filling. For the part isolation, we present a robust and efficient method which identifies clusters of line segments belonging to the same part, and computes the rectangular region enclosing that part. In the grasp planning, we use the simple method for determining the occlusion relations, topmost parts, and grasp point. In using the frame coherence, we first identify the bounding window enclosing the changed portion of the successive images and restrict all the vision processing within the bounding window.
The experiment is performed with various types of parts. The isolated parts, occlusion relations, topmost parts, and grasp point are correctly identified in the large number of test images. The frame coherence reduces the processing time drastically.
The techniques of the multiple light sources and gap filling are also applicable to the broad class of images other than the bin images for extracting the reliable edges. Our frame coherence idea can easily extended to other application domains of the computer vision that deal with the locally changing objects.