Yolov3 and kalman filter based star detecting algorithm for cubesats' star tracker sensors

High accuracy of the attitude estimation in CubeSats is crucial for the reliability of the system and the complete fulfillment of CubeSat's missions. Star tracker cameras are reliable options for providing highly accurate attitude estimation on the satellites. The position information of the stars is one of the critical issues for high accuracy. Various star detection and position detection algorithms are used. Especially with the development of machine learning, in addition to traditional methods, state-of-the-art algorithms have begun to be developed. The accuracy of star detection in star detection algorithms using machine learning in the literature remains lower than in traditional methods. Therefore, this thesis uses the YOLOv3 deep learning network for star detection as a state of art method, and the traditional Kalman filter to track stars' positions and further increase accuracy by avoiding potential errors. The YOLOv3 algorithm, which has high sensitivity and fast working structure within the real-time object detection algorithms, processes the images from the star tracker cameras, increasing the sensitivity of the attitude estimation for star detection and position information by almost 99 percent. Moreover, in the attitude estimation algorithm developed with the star tracker for the cube satellites, the information coming from the YOLO network is filtered with the Kalman filter, which is one of the most reliable fractionation methods known, and the accuracy rate reaches almost 99 percent.