Abstract: In tasks such as space warning, evasion, and surveillance of noncooperative targets, the accurate acquisition of detailed information about targets requires simultaneous estimation of both their motion state and shape characteristics. Therefore, research on extended target tracking algorithms is critical. To address this situation, a novel algorithm of extended target tracking suitable for three-dimensional orbital space is proposed. Firstly, a non-parametric modeling approach based on Gaussian process (GP) radial functions is used to model three-dimensional shapes, effectively solving the problem where random matrix models fail to describe complex shapes accurately. Secondly, a probability hypothesis density (PHD) multitarget tracking filter based on the random finite set (RFS) theory is explored. The RFS theory is employed to leverage its benefits, including the elimination of explicit data association, and to effectively handle high-density clutter in space. Finally, a dynamic threshold partitioning strategy based on an improved Euclidean distance is proposed. This strategy which significantly enhances computational efficiency while ensuring tracking accuracy. The simulation results demonstrate that, compared with the extended object tracking algorithm based on the random matrix method, the proposed GP-PHD filter exhibits significant improvements in both target state estimation accuracy and 3D shape description. In terms of shape description, the IOU metric demonstrates an enhancement of 64%. This method effectively overcomes the limitations of traditional tracking methods in orbital space and provides a new technical solution for noncooperative target tracking in space.

Key words: space multi-target tracking, Gaussian process, extended target tracking, PHD filter, random finite set