As an emerging field, unmanned aerial vehicles (UAVs) feature from interdisciplinary techniques in science, engineering and industrial sectors. The massive applications span from remote sensing, precision agriculture, marine inspection, coast guarding, environmental monitoring, natural resources monitoring, e.g. forest, land and river, and disaster assessment, to smart city, intelligent transportation and logistics and delivery.
With the fast growing demands from a wide range of application sectors, there is always a bottleneck how to improve the efficiency and efficacy of UAV in operation. Often, smart decision making is needed from the captured footages in a real-time manner, yet this is severely affected by the poor image quality, ineffective object detection and recognition models, and lack of robust and light models for supporting the edge computing and real deployment.
In this thesis, several innovative works have been focused and developed to tackle some of the above issues. First of all, considering the quality requirements of the UAV images, various approaches and models have been proposed, yet they focus on different aspects and produce inconsistent results. As such, the work in this thesis has been categorised into denoising and dehazing focused, followed by comprehensive evaluation in terms of both qualitative and quantitative assessment. These will provide valuable insights and useful guidance to help the end user and research community.
For fast and effective object detection and recognition, deep learning based models, especially the YOLO series, are popularly used. However, taking the YOLOv7 as the baseline, the performance is very much affected by a few factors, such as the low quality of the UAV images and the high-level of demanding of resources, leading to unsatisfactory performance in accuracy and processing speed. As a result, three major improvements, namely transformer, CIoULoss and the GhostBottleneck module, are introduced in this work to improve feature extraction, decision making in detection and recognition, and running efficiency. Comprehensive experiments on both publicly available and self-collected datasets have validated the efficiency and efficacy of the proposed algorithm.
In addition, to facilitate the real deployment such as edge computing scenarios, embedded implementation of the key algorithm modules is introduced. These include the creative implementation on the Xavier NX platform, in comparison to the standard workstation settings with the NVIDIA GPUs. As a result, it has demonstrated promising results with improved performance in reduced resources consumption of the CPU/GPU usage and enhanced frame rate of real-time processing to benefit the real-time deployment with the uncompromised edge computing.
Through these innovative investigation and development, a better understanding has been established on key challenges associated with UAV and Simultaneous Localisation and Mapping (SLAM) based applications, and possible solutions are presented.
Keywords: Unmanned aerial vehicles (UAV); Simultaneous Localisation and Mapping (SLAM); denoising; dehazing; object detection; object recognition; deep learning; YOLOv7; transformer; GhostBottleneck; scene matching; embedded implementation; Xavier NX; edge computing.
Date of Award | 29 Feb 2024 |
---|
Original language | English |
---|
Awarding Institution | - University Of Strathclyde
|
---|
Supervisor | Hong Yue (Supervisor) & Malcolm Macdonald (Supervisor) |
---|