TL;DR
This paper introduces a novel off-grid compressive sensing (CS) approach for simultaneously locating multiple passive sources using a single moving array. By leveraging the sparse representation of array covariance matrices and an iterative majorization-minimization technique with an atan-sum objective function, the method achieves higher localization accuracy with fewer measurements and array elements compared to traditional techniques, even in low SNR conditions.
What Problem Does This Paper Address?
Accurately pinpointing the locations of multiple passive sources (like radar or sonar targets) is a long-standing challenge, especially when using a single moving sensor platform. Traditional methods often rely on two-step processes (e.g., estimating direction of arrival first, then position) which suffer from nonlinearities and poor signal-to-noise ratio (SNR) performance. Furthermore, these methods often assume sources lie on predefined grids, limiting precision and requiring extensive computational resources for finer grids. The problem is exacerbated by the need to uniquely assign measurements to individual targets and the inherent limitations of signal acquisition from non-cooperative receivers. This paper aims to overcome these limitations by proposing a direct, one-step localization method that can handle multiple targets without requiring them to be on fixed grids, and is more robust to noise and limited data.
Methods
The paper proposes a two-stage approach based on compressive sensing (CS):
- On-Grid CS Localization: This initial stage uses an over-complete dictionary of steering vectors from predefined search grids. The problem is formulated as an $\ell_1$-norm minimization problem, solved using convex optimization, to estimate target positions. This provides a baseline localization.
- Off-Grid CS Localization with Iterative Reweighted Majorization-Minimization (MM): To achieve finer precision without dramatically increasing grid density (and thus computational cost), the authors introduce an off-grid approach. This method uses an "atan-sum" objective function, which is shown to be more similar to $\ell_0$-norm and more sparsity-encouraging than log-sum functions. The core innovation here is the use of a majorization-minimization technique to iteratively replace the non-convex atan-sum objective with a quadratic convex function that can be easily minimized. This iterative process refines both the target positions and the dictionary itself, effectively learning the optimal locations without relying on pre-defined grids. The paper also proves the equivalence between the $\ell_0$-norm and the proposed atan-sum objective function under certain conditions.
Datasets
Not discussed in the paper. The paper relies on simulated data for its experiments.
Key Insights and Ideas
- Off-grid is crucial for precision: The paper strongly advocates for off-grid methods to achieve high-resolution localization, as real-world targets rarely align perfectly with predefined grids.
- Atan-sum as a better $\ell_0$ surrogate: The use of the atan-sum function is highlighted as a more effective approximation of the $\ell_0$-norm compared to other common surrogates like log-sum, leading to improved sparsity and robustness.
- MM for tractability: The majorization-minimization technique is key to making the optimization of the non-convex atan-sum objective tractable by iteratively approximating it with convex functions.
- Moving array benefits: The method effectively leverages the "virtual aperture" created by a moving array, allowing for better localization with fewer discrete observing positions.
- Robustness to low SNR: The proposed off-grid approach demonstrates improved performance in low SNR scenarios compared to traditional methods.
Key Results
- The proposed off-grid algorithm significantly outperforms existing methods like JSRACM and SDF in terms of Root Mean Square Error (RMSE), especially at lower SNRs.
- Localization precision is shown to be approximately proportional to the virtual aperture created by the moving array.
- The algorithm requires fewer observing points (M) and array elements (L) to achieve comparable or better performance than traditional methods, making it more practical for scenarios with limited data or sensor constraints. For example, the paper states it only needs three observing points for acceptable performance.
- The method is capable of localizing multiple targets, even when the number of targets exceeds the number of antennas, demonstrating its robustness.
Limitations and Future Work
- Local Minima: The authors acknowledge that the MM-based algorithm may converge to a local minimum due to the non-convex nature of the objective function. They suggest that proper selection of the regularization parameter ($\lambda$) and the initial set of target positions ($S^{(0)}$) can help mitigate this.
- Initialization: The paper proposes initializing the localization area by dividing it into coarse grids to reduce computational complexity and prevent the algorithm from falling into unpredictable local minima.
- Pruning: Computational complexity can be further reduced by pruning (removing) signal components with very small estimated powers at each iteration.
- Future Work: While not explicitly detailed, the paper's focus on improving localization accuracy and robustness with limited data suggests future work could explore real-world deployment challenges, adaptive grid refinement strategies, and integration with other sensor modalities.
Summary updated