Based the fringe projection profilometry, a compact and flexible positionable measuring head can be combined with optical fiber bundles to perform in-situ inspection tasks in industrial applications. Surfaces of complex geometries can be reconstructed and quantified in metric coordinates by means of a fast, non-contact and high-resolution measurement. Defect segmentation, on the other hand, is rather complex with three-dimensional point clouds, since reference data is required or a deviation determination is ambiguous and susceptible to errors. Due to each reconstructed object point corresponding to a camera pixel, it is possible to apply image processing algorithms or neural networks for defect segmentation. Since image based segmentation is more susceptible to poor illumination and deviating surface curvature or texture, a circular array of miniature LEDs has been coaxially arranged around the imaging optics of the camera’s fiber to provide different illumination directions. By utilizing a directional variable illumination sequence, the advantages of image-based segmentation can be combined with the unambiguousness and metric quantifiability of point cloud data.
A 3D measuring endoscope with a small measuring head and parallel arrangement of the fibers can be guided into forming plants and carry out precise measurements of geometries which are unreachable for most three-dimensional measuring systems. The data obtained can be used to quantify the wear of highly stressed structures and thus provide information for maintenance. Due to the compact sensor design and the required accuracy, optics with small working distance and a small measuring volume are used. In addition to in situ single measurements of highly stressed structures, over a hundred individual measurements are conceivable in order to convert large and complex geometries into point clouds. Besides the robust and accurate registration of all measurements, merging is one of the main causes of inaccurate measurement results. Conventional merging algorithms merge all points within a voxel into a single point. Due to the large overlap areas required for registration, points of diverse quality are averaged. In order to perform an improved adaptive merging, it is necessary to define metrics that robustly identify only the good points in the overlapping areas. On the one hand, the 2D camera sensor data can be used to estimate signal-based the quality of each point measured. Furthermore, the 3D features from the camera and projector calibration can evaluate the calibration of a triangulated point. Finally, the uniformity of the point cloud can also be used as a metric. Multiple measurements on features of a calibrated microcontour standard were used to determine which metrics provide the best possible merging.