RMMNet: Deep Memory Aided Extended Object Tracking via High-Order Markovian Modeling and State Decoupling

Grounded in Bayesian filtering (BF), random matrix-based extended object tracking (EOT) offers an efficient framework for the joint estimation of a target’s kinematic state and extension. However, the performance of these methods is fundamentally constrained by the first-order Markov assumption, which often fails to capture the high-order Markovian dynamics prevalent in real-world scenarios. To overcome this limitation, we introduce RMMNet, a novel EOT method that integrates a deep memoryaided recursive neural network filter within a BF framework, enabling robust tracking of targets that exhibit high-order Markovian dynamics. Initially, we propose an approximate model under a high-order Markovian assumption and derive the implementation of its BF framework. Thereafter, Gaussian approximation and moment matching are employed to derive the analytical formulations for the proposed BF framework. Finally, based on the closed-form formulations, we design the architecture of RMMNet as an end-to-end trainable recursive neural networkbased filter. Experimental results on both simulated and real-world data demonstrate that our method outperforms traditional EOT approaches and other state-of-the-art deep learning methods.