Zhang, Zhongyi Michael
(2025)
A data fusion pipeline for registering point clouds with novel characteristics: enabling the computer to recognise a pattern without training dataset.
PhD thesis, University of Nottingham.
Abstract
Data fusion is the technical process which can provide comprehensive information about an object by combining multiple datasets that are collected by different sensors. It has been employed for point cloud registration in the context of optical coordinate measurement, an important subject in metrology. Researchers in this field have proposed numerous methods to improve the performance of data fusion, which can be categorised into user-dependent methods, including Gaussian process (GP) and weighted least-squares (WLS) algorithms, and user-independent methods such as machine learning. Recent research has shown the convenience of deploying GP and WLS and the flexibility and autonomous functionality of machine learning solutions. However, the target scenarios have been focused on point clouds in similar sizes and point densities. This trend leaves room for further innovation in point cloud registration.
In this research project, a new algorithmic pipeline, which is capable of registering two point clouds with the following characteristics contained in a maximum working volume of 500×500×500 mm, is proposed: 1) the two point clouds are collected from an engineered object via two separate optical measurement systems, i.e. they are located in two uncorrelated coordinate frames; 2) the smaller point cloud shows the surface texture on a small area on the engineered object, which is represented by the larger point cloud; 3) the point density of the smaller point cloud is > 10 times the point density of the larger point cloud. The challenge lies in the omission of training data: the variation of surface texture is infinite and the area on the engineered object cannot be rigorously determined by the user. As such, the thesis proposes a statistical method to register two point clouds with aforementioned characteristics, which can be summarised as the “geometrical similarity comparison”. In the step of detecting the target area, the larger point cloud is subdivided into equally sized subsections (sub-clouds); the geometrical similarity between each sub-cloud and the smaller point cloud is measured via principal component analysis (PCA). The comparison based on PCA will result in the smaller point cloud being located in the target area formed by the selected sub-clouds. Afterwards, the space mutually occupied by the target area and the smaller point cloud is voxelised so that the spatial point distributions of both point clouds can be assessed. The orientation of the smaller point cloud which aligns it to the target area is determined as the correct orientation, and hence completes the whole registration process.
To test the performance of this algorithmic pipeline, three experimental cases were designed with a gradation of geometrical complexity: two cases include synthetic point clouds generated from CAD models and one case in which the point clouds were collected from a coin. The differences of point densities between the pair of point clouds in these cases are in the range of 10 to 10^2. The results indicate that, though manual double-check is needed as the geometries of the test object increases, the algorithmic pipeline is capable of detecting the location in the larger point cloud to register the smaller point cloud. When scanning the point clouds collected from the coin, the most geometrically complex engineered artefact in this research, the pipeline detected the top 0.73% sub-clouds (186 out of 25,459) which potentially formed the target area in the larger point cloud to register the smaller point cloud.
The pipeline is adequate of detecting the most optimised orientation to register the smaller point cloud regardless of the geometrical complexity of both point clouds. With a 10 interval for 360 orientation attempts, the optimal orientation achieves a registration error ranging from 5% to 0.15%, based on the mean point-to-point distance relative to the smaller point cloud’s dimensions.
The directions left for future work are increasing the accuracy of the detection of the target area and renovating the algorithm to reduce the computational cost. Additionally, big-picture topics questioning the empirical methodology of this research are discussed and concluded.
| Item Type: |
Thesis (University of Nottingham only)
(PhD)
|
| Supervisors: |
Piano, Samanta
Catalucci, Sofia
Thompson, Adam
Leach, Richard |
| Keywords: |
Data fusion, point cloud registration, algorithm development, pattern recognition, advanced manufacturing technology |
| Subjects: |
T Technology > TS Manufactures |
| Faculties/Schools: |
UK Campuses > Faculty of Engineering > Department of Mechanical, Materials and Manufacturing Engineering |
| Item ID: |
80234 |
| Depositing User: |
Zhang, Zhongyi
|
| Date Deposited: |
29 Jul 2025 04:40 |
| Last Modified: |
29 Jul 2025 04:40 |
| URI: |
https://eprints.nottingham.ac.uk/id/eprint/80234 |
Actions (Archive Staff Only)
 |
Edit View |
Tools
Tools
