Thesis

Robust perception and detection systems for micro aerial vehicle based intelligent visual inspection in complex environments

Downloadable Content

Download PDF
Creator
Rights statement
Awarding institution
  • University of Strathclyde
Date of award
  • 2023
Thesis identifier
  • T16789
Person Identifier (Local)
  • 201882834
Qualification Level
Qualification Name
Department, School or Faculty
Abstract
  • Structural failures caused by cracks or corrosion lead to catastrophic consequences in environmental, human and economic terms. Therefore, structural health assessments are essential for maintaining their structural integrity. Considering the small size and manoeuvrability of the Unmanned Aerial Vehicles (UAVs), the aerial inspection platform provides an efficient solution for inspecting high-risk sites such as drilling rigs and pressure vessels, which are traditionally inspected by experienced human engineers that mainly rely on their naked eyes. By deploying an autonomous aerial vision-based visual inspection system, the limitations of the human cost and safety factors of previously time-consuming tasks have the potential to be overcome. However, the maturity level of autonomous inspection UAVs still needs to be improved. Motivated by the observations derived from improving the autonomous capability of aerial visual inspection, this thesis presents novel solutions to contribute to autonomous UAVs for asset visual inspection. First and foremost, the feasibility of using a UAV system with Visual Simultaneous Localisation and Mapping (VSLAM) for autonomous visual inspection in confined and low-illumination indoor environments is verified in the simulation environment for the first time. With image contrast-enhanced VSLAM, the UAV can track the planned trajectory stably and record videos. Subsequently, corrosion detection and UAV localisation systems are further investigated to address the challenges that arise when implementing the UAV in complex environments and deploying algorithms on the UAV onboard platform. In particular, to address the computational challenges of implementing a deep learning-based corrosion detector on UAV onboard platforms caused by the extensive usage of traditional convolutional layers, a solution with a lightweight model design is provided, achieving the first UAV onboard deep learning-based real-time corrosion detector. This advancement was achieved through lightweight convolution utilising Depthwise Separable convolution (DSconv), innovative feature extraction and fusion techniques leveraging the Convolutional Block Attention Module (CBAM) and the proposed improved Spatial Pyramid Pooling (SPP), refined detection strategies incorporating three-scale detection, and an optimised learning approach using the focal loss. The proposed lightweight but powerful corrosion detector is verified by leveraging the Nvidia Jetson TX2, and it achieves 20.18 Frames Per Second (FPS) and 84.96% mean Average Precision (mAP). The overall performance meets the requirements and outperforms other state-of-theart detectors. Then, the issue of the degraded performance of VSLAM-based UAV localisation systems in complex lighting and textureless environments is investigated. Initially, the inherent challenges faced by feature-based VSLAM in low-contrast environments, where extracting sufficient feature points is challenging, need to be addressed. To mitigate this issue, adaptive adjustments to the Features from Accelerated Segment Test (FAST) threshold and image enhancement from contrast and sharpening perspectives are proposed. These improvements are then seamlessly integrated into monocular ORB-SLAM3, ensuring a stable and robust extraction of feature points. Compared with other advanced works, the developed VSLAM system achieves overall higher localisation accuracy and robustness in low-contrast environments while maintaining good performance in general environments. To address the performance degradation or failure of VSLAM and Visual Odometry (VO) systems in environments with textureless and low-illumination conditions where sufficient feature points cannot be extracted, the deep learning based feature point extraction method with a novel lightweight model has been investigated and incorporated into a VO system. Specifically, this model has been achieved by incorporating DSconv and Deformable Convolution (DFconv), whose kernel offsets are calculated through DSconv. Extensive experiments, including physical UAV flying tests, have been conducted to validate the feasibility and exceptional performance of the proposed method. Moreover, the developed model allows the UAV to localise itself and track the predefined trajectory in the textureless and challenging lighting environment, where both the other traditional and deep learning involved VO and VSLAM systems fail.
Advisor / supervisor
  • Yang, Erfu
Resource Type
DOI

Relations

Items

Thumbnail Title Date Uploaded Visibility Actions
file details PDF of thesis T16789 2023-12-15 Public Download