Three-Dimensional Multiple Extended Targets Tracking Under Occlusion Using Variational Gaussian Processes

With advances in high-resolution sensor technology, three-dimensional (3D) multi-extended target tracking (METT) has recently gained increasing attention in both civilian and military fields. To track 3D multi-extended targets under occlusion, a variational Gaussian process METT (VGP-METT) algorithm is proposed in this article. First, a deviation augmentation state (DAS) of the multi-target attitude state is proposed, which incorporates the displacement and the velocity deviation between targets. This representation uses information from the unoccluded target to reconstruct the occluded target state, and it increases the target track continuity. Second, based on the DAS model, an association latent variable is introduced to address the data association problem in multi-target tracking, and the occluded target state is extracted from the DAS model. In contrast to traditional measurement partition approaches which are based on distance or density, this variable is jointly modeled with the target states to eliminate the requirement for a priori knowledge of the target. Finally, the VGP-based 3D METT is proposed to track multi-extended targets under occlusion. We utilize the variational Bayesian technique for approximate inference to solve the intractable posterior distribution caused by the introduction of a latent variable. The effectiveness of the proposed algorithm is assessed through both simulations and real-world scenarios.