Novel Multi-Target Tracking Method: PMBM Filter Combined SVD-SCKF with GP-Driven Measurements

Area of Science:

• * Signal Processing
• * Estimation Theory
• * Data Fusion

Background:

• * Existing Poisson multi-Bernoulli mixture (PMBM) recursions struggle with unstable covariance propagation under nonlinear dynamics and measurement association uncertainty.
• * High clutter density and nonlinearities in multi-target tracking scenarios pose significant challenges for traditional filtering methods.

Purpose of the Study:

• * To develop a numerically stable and robust multi-target tracking filter for nonlinear and high-clutter environments.
• * To enhance the prediction and update stages of the PMBM recursion to handle model mismatch and clutter effectively.

Main Methods:

• * Implemented singular value decomposition (SVD) within the square-root cubature Kalman filter (SCKF) for stable covariance propagation.
• * Integrated SVD-SCKF into the PMBM prediction step for propagating Gaussian-mixture components.
• * Utilized Gaussian processes (GP) to regress unknown measurement functions and adaptively determine gating and target discrimination.

Main Results:

• * The proposed SVD-SCKF-enhanced PMBM filter demonstrates improved numerical stability and robustness in predictions under nonlinear dynamics.
• * Adaptive gating and Bayesian target discrimination effectively reduce clutter impact and improve measurement-to-target association.
• * Simulations confirm the filter's effectiveness and stability in complex multi-target tracking scenarios.

Conclusions:

• * The novel filter significantly enhances the performance of PMBM recursion in challenging tracking conditions.
• * The combination of SVD-SCKF and GP regression offers a powerful approach for robust multi-target tracking.
• * This work provides a stable and effective solution for real-world applications demanding accurate multi-target tracking.