Can Machine Learning be Used as a Tool to Clean 3D Point Clouds for CAD Model Construction & Meshing?

点击翻译按钮获取中文摘要

This project addresses the problem of cleaning point clouds as a pre-processing step for creating respective CAD models and meshes, ultimately for numerical engineering applications. Methods of visualising the findings will also be explored.High resolution laser scans rarely come without noise or imperfections. These scans typically come in the form of point clouds, high resolution arrangements of points which represent surfaces. When considering engineering applications, these three-dimensional point clouds are subsequently transformed into a mesh as part of the reverse-engineering process, whereby the points are connected via the surfaces of interconnected polygons. Simulations can then be ran using these meshes and analysed using methods such as finite element analysis (FEA). The goal of cleaning these scans is to remove the points which are either not relevant for the purpose or counter-productive. Any noise or imperfections in the original scan can be exacerbated when the cloud is converted in to a mesh which will have consequences when running simulations on the model. Manually cleaning these scans can be a very time-intensive process and while many software packages offer the ability to do this, an efficient automated or semi-automated process would be beneficial. In this project, a machine learning approach is adopted to pre-process point clouds by considering the subsequent steps in the reverse-engineering and numerical engineering pipeline. The focus will initially be on classification and clustering algorithms to find and identify imperfections which could affect the analysis results in later stages.It has been previously shown that boosted random forests can be used to aid a user in point-cleaning. One investigation will be to identify if more complex machine learning algorithms can further automate this process without a drastic increase in inefficiencies such as time and computer resources. Deep learning is an exciting, relatively modern area which is showing positive applications in many areas and so, ultimately, will be explored for the aforementioned implementation.An interesting part of this project will be investigating the effects of various input features for the models. Cartesian-based (or potentially otherwise) locations, respective normals, RGB colour channels and light intensities are some of the parameters which will be considered. Variations in the parameters in the models themselves will also be tested; numbers of decision trees in forests, pooling methods and numbers of layers in neural networks, variations in k in k-clustering algorithms are examples of this. Dimensionality reduction techniques such as principal component analysis will also be investigated to see if increases in performance or efficiency can be achieved.Visualisation of the results of these processes is an important part of the later stages of this project. Good visualisation of the work done in relevant areas of the scans could provide benefits to those potentially working in manually further cleaning the scans as well as engineers working directly at the meshing and numerical-engineering stages. Comparisons of the methods of visualising the various stages of this processing will be conducted.