Use this URL to cite or link to this record in EThOS: https://ethos.bl.uk/OrderDetails.do?uin=uk.bl.ethos.785124
Title: Machine learning techniques for 3D data analysis
Author: Duan, Jifang
ISNI:       0000 0004 7970 6662
Awarding Body: UCL (University College London)
Current Institution: University College London (University of London)
Date of Award: 2019
Availability of Full Text:
Access from EThOS:
Full text unavailable from EThOS. Restricted access.
Access from Institution:
Abstract:
The aim of the research in this thesis is to investigate and develop machine learning classification and recognition techniques for application in the field of 3D data analysis. The research focuses on three topics using two different types of data: 3D force touch panel display data for user touch behaviour classification, Light Detection and Ranging (LiDAR) 3D point cloud data for sphere recognition and for complex 3D object segmentation and recognition. In the research on 3D force touch detection data on piezoelectric-based interactive panel displays, a nested structured deep neural network is proposed to provide customised touch behaviour analysis. This is a new application of deep learning technology in piezoelectric-based 3D force sensing. In the research on sphere recognition, artificial neural network-based and Range-based novel recognition algorithms are proposed. The Range method is proved to be time efficient and memory saving compared to conventional methods, whereas the ANN-based method is found not to be optimal for dealing with the sphere recognition problem as this problem can be easily simplified into pure mathematical models. In the research on 3D object segmentation and recognition, an automated approach that takes the raw LiDAR scanning data as input and outputs the recognised 3D objects is proposed. This method consists of the following three steps: plane detection, object segmentation, and object recognition. In order to reduce the processing time and the computational complexity of the algorithm to make it possible to run in a reasonable time on an individual processing device with limited memory and computing resources, a novel plane detection approach is proposed. A fast voxel grid-base region growing method is applied to object segmentation. Novel convolutional neural network architectures that take volumetric representation of 3D objects with small voxel grid sizes as inputs are also implemented in the system.
Supervisor: Not available Sponsor: Not available
Qualification Name: Thesis (Ph.D.) Qualification Level: Doctoral
EThOS ID: uk.bl.ethos.785124  DOI: Not available
Share: