Transmission tower inspection is a crucial step in maintaining high-voltage electrical infrastructure. To perform these inspections, UAVs have proven to be a groundbreaking method for conducting such inspections. The UAV provide a fast, precise, cheap, and modular way to perform thorough audits. The next step in the transmission tower audit is to automate the analysis task. During this thesis, the goal of achieving a relative and adaptive autonomous flight for the transmission tower was attempted.
In the first part of the thesis, we explain how the audits are executed. We explain the history of the UAV and describe them. We then make a small state-of-the-art Computer Vision neural network. This first part brings the basic understanding of the domain.
In the second part of the thesis, we present our distinct contributions. Due to the sponsorship of this thesis, we highlight the industrial and scientific contributions.
The first contribution is the creation of a platform that enables a comprehensive analysis of flights within the domain. That platform displayed flows and offered some guidance on how to perform audits more effectively. Thanks to the software, we were able to showcase the capabilities of our autonomous system in comparison to other solutions.
The second scientific contribution is the creation of datasets around the domain of transmission towers. To train the machine learning algorithm, well-curated datasets were necessary. Due to the scarcity of data in that domain, we decided to study the impact of synthetic data to help train AI in computer vision. The datasets showed promising results in different ML algorithms such as UNet and Mask2Former. In some situations, hybrid dataset-trained models outperform models trained on only the physical world dataset.
The third scientific contribution involves training a Convolutional Neural Network for transmission tower segmentation. This modern neural network is designed for use in an embedded computer on a UAV.
The fourth contribution is the development of different flight controller algorithms during the thesis. These autopilots ranged from a more basic flight assistant for the pilot to an autonomous flight system. A large-scale comparison is done of the different flight controllers.
The final scientific contribution is a study on monocular depth perception in the context of a UAV flying around a transmission tower. This study compares six state-of-the-art pre-trained models. A large data set is created using photogrammetry software to generate pixel-wise depth annotations for each image. The study compares the models using different metrics and ranks them based on their capabilities.
