Centroid algorithm for high-dynamic star sensor based on key point detection deep learning algorithms

Under dynamic conditions, star spots will move on the image plane of the star sensor, resulting in trailing of the star map. This trailing can significantly reduce the accuracy of star centroid positioning, thereby affecting satellite attitude determination. Unlike traditional methods that restore blurred star maps before positioning, we treat the centroid of the star point as a key point in the trailing star map and use a deep learning model based on object detection to convert the positioning of star points under dynamic conditions into the positioning of key points in the trailing star map. Compared with several traditional centroid algorithm under dynamic conditions, our algorithm not only improves the accuracy of centroid positioning by more than one order of magnitude, but also addresses issues that are unresolved by traditional methods, such as the positioning of non-uniform and non-linear star trailing tracks, positioning under trailing star streak overlap conditions, and demonstrates high robustness to noise. This work introduces what we believe to be a new paradigm for centroid positioning in dynamically blurred star images, pushing the upper limit of centroid positioning accuracy in high-dynamic scenarios and providing a solution for precise attitude determination of star sensors in complex scenarios.