A CNN-based architecture for relative state and target motion parameters estimation in active debris removal missions

This paper presents an original relative navigation architecture for proximity operations towards an uncooperative inactive satellite, employing convolutional neural networks for monocular-based pose estimation. In particular, this latter task is performed following an indirect approach, in which two neural networks of the YOLO family are employed to respectively locate the target satellite in the image and identify a set of key-points corresponding to natural target features. The resulting correspondences between the 3D coordinates of the key-points and their positions in the image are used to solve a Perspective-n-Point problem combining analytical and numerical solvers. In addition, an original strategy is introduced to get estimates of the target angular velocity and of its covariance. Such pose and target angular velocity information are integrated as measurements in the correction step of a Multiplicative Extended Kalman Filter at different frequencies; this approach allows improving rotational state estimation accuracy, especially when dealing with tumbling satellites. A new dataset of 20000 synthetic images of the ENVISAT satellite is generated using the open-source software Blender, accounting for variable conditions in terms of target relative position, attitude, Sun illumination and presence of the Earth in the background, to train the networks. Several testing datasets are generated to validate the proposed methods, simulating monitoring and approach scenarios with spinning and tumbling targets. A parametrical study is also conducted to assess the effect of the size of the key-points’ extraction network and of the suppression of the target detection network on relative navigation accuracy. Results show the capability to estimate the relative state with errors of less than 6 % of the range on relative position and with degree-level accuracy on relative attitude, while relative velocity and target angular velocity are estimated with errors of less than 6 cm/s and 0.1°/s, respectively; the accuracy of target angular velocity estimation is preserved even with tumbling satellites having angular rates up to 1°/s.