Advanced motion estimations and predictions of a tumbling, non-cooperative space object during long-term occlusion

This study aims to improve the rotational motion and inertia parameters estimation performance of an Unscented Kalman Filter model (UKF) for a torque-free tumbling non-cooperative space object using the Gaussian Process (GP). The traditional UKF algorithm which is a physics-based estimation algorithm for non-linear systems is susceptible to the physical process, measurement sampling rate, and filter design. Consequently, slight inaccuracy in the assumed physical models, low sampling rates, or small variations of the filter parameters can result in poor estimation performance. Additionally, the UKF model might not predict the motion and inertia parameters with good accuracy in the absence of sensor measurements, also known as occlusion, a quite common challenge for space missions. To make a UKF model more robust to the factors above, we utilize multi-output GP models with periodic kernels to make long-term predictions of the position and attitude measurements obtained from a Laser Camera System (LCS). These measurement predictions from GP models are used as the sensor measurements for the UKF model. We implement a Fast Fourier Transform on the sensor measurements to determine the initial guess for periodicity hyper-parameters for the periodic kernels. Results from conducted simulations show that the proposed UKF model with GP-predicted measurements (UKF-GP model) performs remarkably well compared to the UKF model under the assumption of long-term occlusion. It is also observed from the results that, the UKF-GP model is more robust to sensor sampling rate, underlying physical process, and filter parameters even with occlusion, compared to the UKF model without occlusion.