Session
Weekend Poster Session 1
Location
Utah State University, Logan, UT
Abstract
With space becoming more and more crowded, there is a growing demand for increasing satellite lifetimes and performing on-orbit servicing (OOS) at a scale that calls for autonomous missions. Many such missions would require chaser satellites to autonomously execute safe and effective flightpath to dock with a non-cooperative target satellite on orbit. Performing this autonomously requires the chaser to be aware of hazards to route around and safe capture points through time, i.e., by first identifying and tracking key components of the target satellite. State-of-the-art object detection algorithms are effective at detecting such objects on a frame-by-frame basis. However, implementing them on a real-time video feed often results in poor performance at tracking objects over time, making errors which could be easily corrected by rejecting non-physical predictions or by exploiting temporal patterns. On the other hand, dedicated object tracking algorithms can be far too computationally expensive for spaceflight computers. Considering this, the paradigm of tracking-by-detection works by incorporating patterns of prior-frame detections and the corresponding physics in tandem with a base object detector. This paper focuses on comparing the performance of object tracking-by-detection algorithms with a YOLOv8 base object detector: namely, BoTSORT and ByteTrack. These algorithms are hardware-in-the-loop tested for autonomous spacecraft component detection for a simulated tumbling target satellite. This will emulate mission conditions, including motion and lighting, with a focus on operating under spaceflight computational and power limitations, providing an experimental comparison of performance. Results demonstrate lightweight tracking-by-detection can improve the reliability of autonomous vision-based navigation.
SSC23-WP1-03 Poster
Comparison of Tracking-By-Detection Algorithms for Real-Time Satellite Component Tracking
Utah State University, Logan, UT
With space becoming more and more crowded, there is a growing demand for increasing satellite lifetimes and performing on-orbit servicing (OOS) at a scale that calls for autonomous missions. Many such missions would require chaser satellites to autonomously execute safe and effective flightpath to dock with a non-cooperative target satellite on orbit. Performing this autonomously requires the chaser to be aware of hazards to route around and safe capture points through time, i.e., by first identifying and tracking key components of the target satellite. State-of-the-art object detection algorithms are effective at detecting such objects on a frame-by-frame basis. However, implementing them on a real-time video feed often results in poor performance at tracking objects over time, making errors which could be easily corrected by rejecting non-physical predictions or by exploiting temporal patterns. On the other hand, dedicated object tracking algorithms can be far too computationally expensive for spaceflight computers. Considering this, the paradigm of tracking-by-detection works by incorporating patterns of prior-frame detections and the corresponding physics in tandem with a base object detector. This paper focuses on comparing the performance of object tracking-by-detection algorithms with a YOLOv8 base object detector: namely, BoTSORT and ByteTrack. These algorithms are hardware-in-the-loop tested for autonomous spacecraft component detection for a simulated tumbling target satellite. This will emulate mission conditions, including motion and lighting, with a focus on operating under spaceflight computational and power limitations, providing an experimental comparison of performance. Results demonstrate lightweight tracking-by-detection can improve the reliability of autonomous vision-based navigation.
