Work

Görüntü Tabanli Hi̇bri̇t Metotlarin Kullanimi İle Uzay Araci Bağil Yöneli̇m Kesti̇ri̇mi̇

Year: 2026
Type: dissertation
Abstract: Autonomous spacecraft proximity operations require reliable relative pose and motion estimation, particularly for uncooperative, tumbling targets. This thesis investigates monocular vision for relative navigation with explicit modelling of measurement uncertainty, using the existing synthetic SPEED-UE-Cube dataset. The framework combines deep learning for keypoint detection, Perspective-n-Point (PnP) pose estimation, and a Multiplicative Extended Kalman Filter (MEKF). It derives a pose covariance from reprojection geometry, adds robust adaptation of the measurement noise, and investigates learned keypoint uncertainty propagated through the pose solve to provide a measurement covariance that varies between images. These approaches are compared with a fixed covariance and a probabilistic PnP baseline in the same MEKF, considering pose and motion estimation accuracy, uncertainty calibration, and filter consistency. A covariance that varies between images improves the filtered estimate on an identical pose stream and reduces the dependence on explicit outlier handling, while point accuracy and covariance quality prove to be distinct properties that must be evaluated separately.
Language: English
Open Access status: green
OpenAlex ID: W7222378542
Cites: 103
Cited by:
Related to: 10