Extended Object Tracking with Inaccurate Heavy-Tailed Noises

In most extended object tracking (EOT) filters employing random matrix, the accurate Gaussian noise is typically treated as prior knowledge within the recursive framework. However, the Gaussian noise assumption frequently fails in practical scenarios, particularly due to outliers induced by sensor malfunctions, and noise information is often difficult to obtain accurately as well. This paper proposes an adaptive EOT filter for the joint estimation of kinematic and shape states under conditions of inaccurate, heavy-tailed process and measurement noises. First, one-step predicted probability density function (PDF) and likelihood PDF are modeled as Student’s t-inverse Wishart (STIW) distribution with different degrees of freedom, where the kinematic state follows a Student’s t distribution and the shape state follows an inverse Wishart distribution. Second, the STIW distribution is reformulated as a hierarchical model utilizing the Gamma distribution, facilitating its integration into the proposed filter’s recursion. Subsequently, the recursion of adaptive EOT filter is derived based on the variational Bayesian method. Finally, the efficacy of the proposed filter is validated through both simulated scenarios and a real-data experiment. Results demonstrate that the proposed adaptive EOT filter exhibits superior performance and robustness compared to contrast filters under inaccurate heavy-tailed noise conditions.