Neural network stereo image segmentation for directed coordinate measuring machine part programming

Abstract

Reverse engineering is used in industry to create a geometric model from an existing physical part. Often, this task is initiated by the collection of 3-D data using a contact sensor, such as a touch probe, mounted onto the end effector of a coordinate measuring machine (CMM). A CCD camera, mounted along with the touch probe, is utilized to automatically direct the digitization process. Stereo images taken with the CCD camera are used to both segment the object into its component surface patches and to locate the object on the CMM bed. A Kohonen self-organizing network is used to segment the stereo images. Areas of constant grey level intensity are used as seed locations from which patches are grown in the network. The stronger patches (those in areas of constant grey level intensity) compete and eventually dominate neighbouring patches in areas of less grey level consistency. In this winner take all strategy, the number of surface patches need not be known beforehand, and learning phases are unnecessary. This is ideal for reverse engineering, as the object shape and complexity is often unknown. Surface patch location is achieved by matching segmented patches between the stereo image pairs. Experiments on a planar test object demonstrate this system's robustness. By combining a CCD camera with the CMM touch probe, an automated 3-D digitization system is developed.