Analytic-Jacobian-Free Neural Measurement Model Embedded Bayesian Filtering

Target tracking aims to estimate target states from noisy observations. Bayesian filtering methods often rely on an explicitly defined measurement model h(·) or accurate computation of its Jacobian matrix. However, such requirements are often violated in practical scenarios involving noncooperative targets or sensors with partially unknown and uncalibrated parameters. To handle this problem, this paper proposes an Analytic-Jacobian-Free Neural Measurement Model Embedded Bayesian Filtering (AJF-NMBF). Specifically, the proposed method replaces the analytic measurement function with a neural operator termed the Differentiable Neural Measurement Model (DNMM), whose corresponding Jacobian is automatically computed through backpropagation. This design preserves the Bayesian inference structure while enabling the learning of complex and unknown nonlinear measurement relationships. Moreover, we develop a Space-Based Optical Angle Tracking Trajectory dataset (namely, SOATT). Experimental results on both our self-developed SOATT dataset and the publicly available ground-based radar dataset (LAST) demonstrate that the proposed DNMM achieves measurement accuracy comparable to analytic measurement functions. Meanwhile, AJF-NMBF achieves tracking performance comparable to model-based Bayesian filters with analytic measurement models across different noise levels and under both non-switching and switching-motion scenarios.