Cross-Sparsity-Enabled Multipath Perception via Structured Bayesian Inference for Multi-Target Estimation

TL;DR

This work tackles multi-target sensing in multipath environments by treating first-order inter-target reflections, which share angles with direct-path targets, as informative priors rather than interference. It introduces a cross sparsity structure within a three-layer hierarchical (3LHS) prior and develops SF-TVBI, a structured fast turbo variational Bayesian inference algorithm that uses two-timescale updates and efficient message passing to handle the cross-path dependencies. The proposed approach yields substantial gains in angle estimation accuracy, achieving performance close to Turbo-VBI while significantly reducing computational complexity. The method demonstrates improved robustness and practicality for real-time, high-resolution sensing in multipath scenarios.

Abstract

In this paper, we investigate a multi-target sensing system in multipath environment, where inter-target scattering gives rise to first-order reflected paths whose angles of departure (AoDs) and angles of arrival (AoAs) coincide with the direct-path angles of different targets. Unlike other multipath components, these first-order paths carry structural information that can be exploited as additional prior knowledge for target direction estimation. To exploit this property, we construct a sparse representation of the multi-target sensing channel and propose a novel cross sparsity structure under a three-layer hierarchical structured (3LHS) prior model, which leverages the first-order paths to enhance the prior probability of the direct paths and thereby improve the estimation accuracy. Building on this model, we propose a structured fast turbo variational Bayesian inference (SF-TVBI) algorithm, which integrates an efficient message-passing strategy to enable tractable probabilistic exchange within the cross sparsity, and a two-timescale update scheme to reduce the update frequency of the high-dimensional sparse vector. Simulation results demonstrate that leveraging the proposed cross sparsity structure is able to improve the target angle estimation accuracy substantially, and the SF-TVBI algorithm achieves estimation performance comparable to that of the Turbo-VBI, but with lower computational complexity.

Figures (9)

• Figure 1: System model of multi-target sensing under multipath propagation.
• Figure 2: Illustration of cross sparsity.
• Figure 3: Overall framework of the proposed SF-TVBI algorithm.
• Figure 4: Factor graph of $p(\mathbf{s})$ in module B.
• Figure 5: RMSE versus the out-loop iteration number.