Development of a deep learning framework for accelerated training and efficient classification of aerial LiDAR data

The energy industry has operational needs to monitor high value assets and infrastructure such as pipelines and powerlines, against intrusions or anomalies. Aerial LiDAR scanners enable around the clock surveillance, as they are active sensors emitting laser pulses to capture detailed point cloud data of surface objects under any lighting conditions. However, there are challenges in processing the large data volume and varying point density collected from LiDAR surveys. Machine learning is used for semantic segmentation, classifying points into object classes, though the high computational demands usually necessitate post-mission analysis. Classification is challenging since point clouds are three-dimensional and object classification methods must be rotationally invariant. A key step in this process is calculating geometric features which describe local shape, size or orientation characteristics of the point cloud, such as verticality, planarity, density and roughness. As part of this research, the most significant geometric features for asset monitoring are identified and integrated into a new hybrid voxel-based deep learning classification framework. This Modified PVCNN framework processes high-density aerial LiDAR datasets along with selected geometric features, with significantly low training times (e.g. 1 hour), operating on minimal computational hardware specifications yet delivering high accuracy for common object classes. The issues of class imbalance are also addressed; whereby physically small or thin-shaped objects are under-represented in aerial LiDAR training datasets. To overcome this issue, a second classification pass is used to improve the initial prediction for such imbalanced object classes. In the second pass, semantically homogeneous points from the initial classification are grouped prior to computing the geometric features used in the classification stage. By doing so, these engineered geometric features better represent the different object classes, thus improving accuracy for the under-represented classes. Finally, we demonstrate the MPVCNN as part of a change detection processing pipeline for a LiDAR time-series dataset. With the predicted class labels from the classifier, the change detection task can be focused on targeted object types, identifying presence or omission from the point clouds, and thereby automate alerts to possible intrusions or anomalies when monitoring assets.

Grantor dc:publisher

Mathematical and Computer Sciences

Year dc:date.issued

2025

Author dc:creator

• Othman, Fauzy Omar Basheer Bin

Advisors dc:contributor.advisor

• Bartie, Doctor Phil
• Chen , Doctor Dongdong
• Lemon, Professor Oliver

Language dc:language.iso

en

Repository record dc:identifier.uri

https://www.ros.hw.ac.uk/handle/10399/5322

OAI record oai:identifier

https://www.ros.hw.ac.uk/oai/request?verb=GetRecord&metadataPrefix=dim&identifier=oai%3Aros.hw.ac.uk%3A10399%2F5322