Cai, Yuanzhi
(2023)
Image-based Semantic Segmentation of Large-scale Terrestrial Laser Scanning Point Clouds.
PhD thesis, University of Liverpool.
|
Text
201218781_May2023.pdf - Author Accepted Manuscript Download (11MB) | Preview |
Abstract
Large-scale point cloud data acquired using terrestrial laser scanning (TLS) often need to be semantically segmented to support many applications. To this end, various three-dimensional (3D) methods and two-dimensional (i.e., image-based) methods have been developed. For large-scale point cloud data, 3D methods often require extensive computational effort. In contrast, image-based methods are favourable from the perspective of computational efficiency. However, the semantic segmentation accuracy achieved by existing image-based methods is significantly lower than that achieved by 3D methods. On this basis, the aim of this PhD thesis is to improve the accuracy of image-based semantic segmentation methods for TLS point cloud data while maintaining its relatively high efficiency. In this thesis, the optimal combination of commonly used features was first found, and an efficient manual feature selection method was proposed. It was found that existing image-based methods are highly dependent on colour information and do not provide an effective means of representing and utilising geometric features of scenes in images. To address this problem, an image enhancement method was developed to reveal the local geometric features in images derived by the projection of point cloud coordinates. Subsequently, to better utilise neural network models that are pre-trained on three-channel (i.e., RGB) image datasets, a feature extraction method (LC-Net) and a feature selection method (OSTA) were developed to reduce the higher dimension of image-based features to three. Finally, a stacking-based semantic segmentation (SBSS) framework was developed to further improve segmentation accuracy. By integrating SBSS, the dimension-reduction method (i.e. OSTA) and locally enhanced geometric features, a mean Intersection over Union (mIoU) of 76.6% and an Overall Accuracy (OA) of 93.8% were achieved on the Semantic3D (Reduced-8) benchmark. This set the state-of-the-art (SOTA) for the semantic segmentation accuracy of image-based methods and is very close to the SOTA accuracy of 3D method (i.e., 77.8% mIoU and 94.3% OA). Meanwhile, the integrated method took less than 10% of the processing time (52.64s versus 563.6s) of the fastest SOTA 3D method.
| Item Type: | Thesis (PhD) |
|---|---|
| Uncontrolled Keywords: | Deep learning, point cloud, semantic segmentation, terrestrial laser scanning (TLS), ensemble learning, band selection, cloud detection, multispectral images |
| Divisions: | Faculty of Science and Engineering > School of Engineering |
| Depositing User: | Symplectic Admin |
| Date Deposited: | 29 Nov 2023 11:58 |
| Last Modified: | 29 Nov 2023 11:59 |
| DOI: | 10.17638/03173140 |
| Supervisors: |
|
| URI: | https://livrepository.liverpool.ac.uk/id/eprint/3173140 |
Dimensions
Dimensions
