Abstract:We propose a message passing algorithm for tracking of clutter signals in MIMO radar. The method exploits basis expansion to linearise the signal model, to enable mean field approach for tracking the posterior distribution of the clutter as it evolves across time, as well as the mean and precision of the clutter map. The method shows good estimation accuracy in simulations for a scenario that adhere to the statistical model used for derivation as well as one that does not. The complexity of the method is linear in both the amount of parameters chosen and the amount of data under consideration.
Abstract:We propose a distributed joint localization and tracking algorithm using a message passing framework, for multiple-input multiple-output radars. We employ the mean field approach to derive an iterative algorithm. The obtained algorithm features a small communication overhead that scales linearly with the number of radars in the system. The proposed algorithm shows good estimation accuracy in two simulated scenarios even below 0 dB signal to noise ratio. In both cases the ground truth falls within the 95 % confidence interval of the estimated posterior for the majority of the track.
Abstract:In this paper, we propose a direct multiobject tracking (MOT) approach for MIMO-radar signals that operates on raw sensor data via variational message passing (VMP). Unlike classical track-before-detect (TBD) methods, which often rely on simplified likelihood models and exclude nuisance parameters (e.g., object amplitudes, noise variance), our method adopts a superimposed signal model and employs a mean-field approximation to jointly estimate both object existence and object states. By considering correlations within in the radar signal due to closely spaced objects and jointly estimating nuisance parameters, the proposed method achieves robust performance for close-by objects and in low-signal-to-noise ratio (SNR) regimes. Our numerical evaluation based on MIMO-radar signals demonstrate that our VMP-based direct-MOT method outperforms a detect-then-track (DTT) pipeline comprising a super-resolution sparse Bayesian learning (SBL)-based estimation stage followed by classical MOT using global nearest neighbour data association and a Kalman filter.