Leveraging light-curve inversion for real-time kinematic state estimation of uncooperative targets

The rapidly increasing number of resident space objects in Earth’s orbit poses a significant threat to the sustainability of current and future space missions. As Active Debris Removal (ADR) and In-Orbit Servicing (IOS) emerge as critical operations for ensuring the safety and longevity of orbital activities, precise and responsive real-time tracking of the kinematic state of non-cooperative target objects becomes a key enabling technology. In this work, we demonstrate we can improve the performance of an Unscented Kalman Filter (UKF) by exploiting readily available light-curves to extract kinematic priors. Attitude and position estimations from a pose estimation algorithm are fed as the primary input to the UKF, with a sampling rate of up to 1 sample/s in the worst case. Specifically, we consider pose estimations from a monocular pose estimation neural network working on resolved images containing targets ranging from 50 m to 4 km. Such estimates, however, can be provided by any other suitable state-of-the-art model, enabling flexibility in the actual implementation. The key novelty of this work is the integration of kinematic motion priors extracted from the photometric analyses of light-curves captured by long-exposure astronomical observatories. This approach leverages light-curve inversion techniques to extract information on the rotation rate and axis of the target object. The results are then provided as an initial condition to the UKF. As orbital debris can be poorly classified, our work does not assume the availability of a CAD model. Performance is assessed by the generation of our own synthetic light-curve dataset including variations in image noise intensity and the degradation in assumed object properties, such as geometry, reflectivity and inertial parameters. Statistics from the machine learning pose estimation results are used to inform the design of the UKF covariance matrices associated with state, process noise and measurement noise. The design phase, including tuning of the filter parameters, is carried out on synthetic data. Validation of the model is performed on high-fidelity images of the VEga Secondary Payload Adapter (VESPA), left in LEO following the 2013 launch of a Vega rocket from ESA’s spaceport in Kourou, French Guiana. This data is accessed in collaboration with ClearSpace. The expected results of the outlined state estimation model with kinematic prior integration are increased responsiveness and reliability, especially considering the mild assumptions required. This novel approach aligns with the increasing emphasis on sustainable orbital practices and ESA’s Zero Debris Charter by addressing crucial challenges in real-time proximity operations. It enhances the feasibility of ADR, IOS, and future In-Orbit Construction missions by enabling accurate and efficient rendezvous operations with tumbling objects, potentially characterized by complex dynamics and lesser-known properties. This research builds upon recent advancements in light-curve analysis and real-time pose estimation, and exemplifies a multi-disciplinary solution leveraging computer vision, observational astronomy and spacecraft control, to help address urgent challenges in space sustainability.