Object detection and classification

Abstract

Ensuring the safety of space missions and promoting the long-term sustainability of orbital operations necessitates efficient detection and avoidance of collisions with space debris. This research investigates the application of various YOLO (You Only Look Once) algorithms, a family of advanced real-time object detection models, for identifying and tracking debris in Earth’s orbit. The methodology involves training these models on a comprehensive dataset containing both debris and non-debris objects. Experimental results confirm that YOLO-based models achieve high levels of accuracy, precision, and recall in detecting and classifying space debris. In addition to conventional object detection, this study integrates segmentation-based techniques that enable pixel-level classification of debris objects. This enhances the granularity of detection by allowing the extraction of detailed shape, size, and positional information. Using camera resolution parameters, pixel-based segmentation data is converted into real-world dimensions, supporting accurate estimation of object area, which is crucial for assessing collision risks and informing mission planning. The proposed detection pipeline is implemented on multiple NVIDIA Jetson platforms—including Jetson Nano, Jetson Orin Nano, and Jetson AGX Orin—with targeted optimizations to reduce inference latency. These enhancements enable real-time debris detection and monitoring in embedded environments. Furthermore, the study incorporates Detectron2, a state-of-the-art framework for instance segmentation, to further refine object localization and discrimination, especially in scenarios involving overlapping or closely situated debris. Beyond its space-focused application, the research demonstrates the versatility of the detection framework by adapting it for use in unmanned aerial vehicle (UAV)-based precision agriculture. The models originally trained for space debris are repurposed to identify and classify crops, pests, and weeds, showcasing the flexibility of advanced vision algorithms across domains. Overall, the findings from this study contribute significantly to the development of intelligent, automated debris monitoring systems that enhance collision avoidance and mission planning. Additionally, the adaptability of these models suggests broad applicability in domains such as autonomous agriculture, environmental surveillance, and resource optimization.