Adaptive YOLO-V8 for Low-Earth Orbit Debris Detection Improving Detection Accuracy with Dynamic Training

Abstract

It is crucial to identify methods to detect debris due to a new tendency that is emerging in LEO orbits, which is threatening the functionality of satellites and interplanetary missions. This research solves this problem using an improved YOLO-V8 model that enables the detection of space debris with higher precision while using adaptive dynamic learning approaches. It was critical for our model to be able to identify and categorize as many kinds of objects as possible, and our database currently contains 11 object classes, including space debris and satellites. To specifically detect small and moving objects, we utilized the YOLO-V8 model, tailoring the train options to the unique object detections in this class. For the training of our model, we used a large quantity of data in addition to the images from these 11 classes and used SGD as our optimizer with the learning rate of 0.25 and individual weight decay parameters. Additionally, we utilized blur and grayscale transforms to enhance the model through data augmentation. By comparing the obtained results, we can observe enhanced detection accuracy in each class separately, as well as a general boost in prediction and recall. Due to the cross-entropy function's flexibility, the model was able to perform well on various object sizes and speeds in an orbital context, making detection consistent. A lot of fine tuning was required in the training parameters in order to get the desired or even better results devoid of false positive detection. This paper describes how YOLO-V8 with adaptive training achieved outstanding results for object and debris detection in low Earth orbit (LEO) to improve space usage safety and define better approaches to space debris management.