• Login
• Admin

Eastwood, Joe (2023) Machine learning for the automation and optimisation of optical coordinate measurement. PhD thesis, University of Nottingham.

Preview

PDF (Including corrections from viva) (Thesis - as examined) - Requires a PDF viewer such as GSview, Xpdf or Adobe Acrobat Reader Available under Licence Creative Commons Attribution. Download (60MB) | Preview

Abstract

Camera based methods for optical coordinate metrology are growing in popularity due to their non-contact probing technique, fast data acquisition time, high point density and high surface coverage. However, these optical approaches are often highly user dependent, have high dependence on accurate system characterisation, and can be slow in processing the raw data acquired during measurement. Machine learning approaches have the potential to remedy the shortcomings of such optical coordinate measurement systems. The aim of this thesis is to remove dependence on the user entirely by enabling full automation and optimisation of optical coordinate measurements for the first time. A novel software pipeline is proposed, built, and evaluated which will enable automated and optimised measurements to be conducted. No such automated and optimised system for performing optical coordinate measurements currently exists. The pipeline can be roughly summarised as follows:

intelligent characterisation -> view planning -> object pose estimation -> automated data acquisition -> optimised reconstruction.

Several novel methods were developed in order to enable the embodiment of this pipeline. Chapter 4 presents an intelligent camera characterisation (the process of determining a mathematical model of the optical system) is performed using a hybrid approach wherein an EfficientNet convolutional neural network provides sub-pixel corrections to feature locations provided by the popular OpenCV library. The proposed characterisation scheme is shown to robustly refine the characterisation result as quantified by a 50 % reduction in the mean residual magnitude. The camera characterisation is performed before measurements are performed and the results are fed as an input to the pipeline. Chapter 5 presents a novel genetic optimisation approach is presented to create an imaging strategy, ie. the positions from which data should be captured relative to part’s specific geometry. This approach exploits the computer aided design (CAD) data of a given part, ensuring any measurement is optimal given a specific target geometry. This view planning approach is shown to give reconstructions with closer agreement to tactile coordinate measurement machine (CMM) results from 18 images compared to unoptimised measurements using 60 images. This view planning algorithm assumes the part is perfectly placed in the centre of the measurement volume so is first adjusted for an arbitrary placement of the part before being used for data acquistion. Chapter 6 presents a generative model for the creation of surface texture data is presented, allowing the generation of synthetic butt realistic datasets for the training of statistical models. The surface texture generated by the proposed model is shown to be quantitatively representative of real focus variation microscope measurements. The model developed in this chapter is used to produce large synthetic but realistic datasets for the training of further statistical models. Chapter 7 presents an autonomous background removal approach is proposed which removes superfluous data from images captured during a measurement. Using images processed by this algorithm to reconstruct a 3D measurement of an object is shown to be effective in reducing data processing times and improving measurement results. Use the proposed background removal on images before reconstruction are shown to benefit from up to a 41 % reduction in data processing times, a reduction in superfluous background points of up to 98 %, an increase in point density on the object surface of up to 10 %, and an improved agreement with CMM as measured by both a reduction in outliers and reduction in the standard deviation of point to mesh distances of up to 51 microns. The background removal algorithm is used to both improve the final reconstruction and within stereo pose estimation. Finally, in Chapter 8, two methods (one monocular and one stereo) for establishing the initial pose of the part to be measured relative to the measurement volume are presented. This is an important step to enabling automation as it allows the user to place the object at an arbitrary location in the measurement volume and for the pipeline to adjust the imaging strategy to account for this placement, enabling the optimised view plan to be carried out without the need for special part fixturing. It is shown that the monocular method can locate a part to within an average of 13 mm and the stereo method can locate apart to within an average of 0.44 mm as evaluated on 240 test images. Pose estimation is used to provide a correction to the view plan for an arbitrary part placement without the need for specialised fixturing or fiducial marking.

This pipeline enables an inexperienced user to place a part anywhere in the measurement volume of a system and, from the part’s associated CAD data, the system will perform an optimal measurement without the need for any user input. Each new method which was developed as part of this pipeline has been validated against real experimental data from current measurement systems and shown to be effective.

In future work given in Section 9.1, a possible hardware integration of the methods developed in this thesis is presented. Although the creation of this hardware is beyond the scope of this thesis.

Item Type:	Thesis (University of Nottingham only) (PhD)
Supervisors:	Piano, Samanta Leach, R. K.
Keywords:	machine learning, metrology, computer vision, 3D scanning, 3D imaging, computational photography, optimisation, automation, calibration, characterisation, photogrammetry
Subjects:	Q Science > Q Science (General) T Technology > TA Engineering (General). Civil engineering (General)
Faculties/Schools:	UK Campuses > Faculty of Engineering
Item ID:	72918
Depositing User:	Eastwood, Joe
Date Deposited:	21 Jul 2023 04:40
Last Modified:	21 Jul 2023 04:40
URI:	https://eprints.nottingham.ac.uk/id/eprint/72918

Actions (Archive Staff Only)

Edit View