Image-based detection and tracking of resident space objects for situational awareness

• Main Author: Zhang, Rangya
• Other Authors: Mir Feroskhan
• Format: Thesis-Master by Research
• Language: English
• Published: Nanyang Technological University 2024
• Subjects: Computer and Information Science; Engineering; Space situational awareness; Space object detection; Space object tracking; Continual learning
• Online Access: https://hdl.handle.net/10356/181510
• Institution: Nanyang Technological University

Space Situational Awareness (SSA) is crucial for ensuring the safety and sustainability of space operations. With the exponential increase in the number of Resident Space Objects (RSOs), including operational satellites, defunct satellites, and space debris, the need for advanced and effective detection and tracking technologies has become more urgent than ever. Existing algorithms often struggle with the unique challenges posed by the space environment, such as low Signal-to-Noise Ratio (SNR), high-velocity objects, small target sizes, and highly variable environmental conditions. These challenges highlight the necessity for more sophisticated methodologies capable of addressing the specific conditions of space. This research addresses these critical needs by developing robust detection and tracking algorithms specifically designed for the space environment and enhancing model generalization across various scenarios. The importance of this work lies in its potential to significantly enhance SSA capabilities, enabling more accurate monitoring of RSOs. By improving detection and tracking technologies, this research aims to support the safe and sustainable use of space, mitigating the risks posed by space debris and ensuring the continued success of space missions. The study is structured into three main work packages (WPs). The first, WP1, focuses on developing advanced detection algorithms using space-based optical images. By adapting state-of-the-art (SOTA) machine learning models such as YOLO, the goal is to overcome the limitations of current detection methods and improve accuracy and robustness in the challenging conditions of space. These adaptations are expected to address issues such as low SNR, high-velocity objects, and small target sizes, which are common in space imagery. The algorithms will be designed to handle the unique characteristics of space environments, ensuring their applicability and reliability. The second, WP2, aims to implement and optimize detection-based multiple object tracking algorithms. These methods will ensure accurate ID assignment and continuity of tracking across consecutive frames, addressing issues related to high object velocities and compact target distributions. By refining the similarity calculation and enhancing the robustness of the tracking algorithms, we aim to achieve high precision in tracking multiple objects simultaneously, even in densely populated space environments. The third, WP3, focuses on incorporating continual learning techniques to enhance the model's generalization capabilities across different space environments and backgrounds. Continual learning will be used to improve adaptability and performance under varying conditions, allowing the detection and tracking algorithms to retain knowledge from previous tasks while adapting to new data. This approach aims to mitigate issues such as catastrophic forgetting and to ensure the long-term applicability of the models in dynamic and evolving space scenarios. Several findings have validated the feasibility of the proposed approaches. Experiments have demonstrated the effectiveness of adapted detection algorithms in handling the low SNR and small target sizes typical of space imagery, leading to significant improvements in detection accuracy and robustness. These results highlight the potential of machine learning adaptations to function effectively under challenging space conditions. Additionally, the tracking algorithms have shown enhanced performance in maintaining accurate tracking across frames, even in the face of inherent space environment challenges, thus minimizing the likelihood of misidentification and tracking loss. Furthermore, the introduction of continual learning has helped to enhance the model's generalization, ensuring stable performance as the models encounter new and diverse space environments. Collectively, these findings suggest that the proposed methodologies are well-suited for long-term, reliable performance in SSA. In summary, this research not only aims to overcome the limitations of current SSA technologies but also contributes to the broader fields of computer vision, machine learning, and continual learning by developing innovative solutions for space object detection and tracking. These advancements are expected to play a critical role in enhancing SSA systems, providing more accurate and reliable monitoring of space activities, and ensuring the long-term safety and sustainability of space operations. By addressing the unique challenges of space and advancing the state of SSA technology, this work will support both scientific exploration and commercial utilization of space.