Non-coherent Sensor Fusion via Entropy Regularized Optimal Mass Transport

Filip Elvander, Isabel Haasler, Andreas Jakobsson, Johan KarlssonICASSP 2019 · OpenReview

TL;DR

This paper introduces a novel method for fusing information from multiple sensor arrays to pinpoint the location of signal sources, even when the arrays are not perfectly aligned or calibrated. By leveraging optimal mass transport and entropy regularization, the approach offers robust and computationally efficient source localization, outperforming traditional methods in scenarios with sensor misalignments.

What Problem Does This Paper Address?

Accurately locating signal sources is crucial in many fields, from radar and sonar to audio processing. Traditional methods often rely on the assumption that sensor arrays are perfectly calibrated, meaning their positions and orientations are precisely known. However, in real-world applications, these assumptions can be violated due to calibration errors, physical disturbances, or synchronization issues. This misalignment can significantly degrade the performance of localization algorithms. While some methods attempt to address this by dividing sensors into subarrays, this paper tackles the problem of fusing information from entirely separate sensor arrays where only aggregated data (like covariance matrices) is available, and cross-correlations between arrays cannot be reliably estimated. This "non-coherent" processing scenario is common in distributed sensor networks where communication bandwidth is limited. The core challenge is to robustly estimate the spatial spectrum (which indicates source locations) from these incomplete and potentially misaligned measurements.

Methods

The paper proposes a method based on Optimal Mass Transport (OMT), a mathematical framework that finds the most efficient way to move "mass" from one distribution to another, minimizing a cost function. Here, OMT is used as a distance metric between spatial spectra. The core idea is to find a spatial spectrum that is a "convex barycenter" of individual spectra estimated from each sensor array.

To make this computationally feasible, especially for higher-dimensional problems, the authors introduce entropy regularization. This technique adds a penalty term to the objective function, which makes the problem strictly convex and allows for efficient iterative solutions. The proposed algorithm is an extension of the popular Sinkhorn iterations, adapted for this specific barycenter problem. This regularization not only improves computational efficiency but also contributes to the method's robustness. The algorithm iteratively refines estimates of the spatial spectrum and associated transport plans, effectively balancing the information from each sensor array while accounting for potential misalignments.

Datasets

Not discussed in the paper. The evaluation is performed using simulated data generated to mimic realistic sensor fusion scenarios.

Key Insights and Ideas

  • OMT as a Robust Distance Metric: Using OMT to measure the "distance" between spatial spectra provides inherent robustness to spatial shifts and distortions caused by sensor misalignments. Unlike point-wise metrics, OMT considers the geometry of the underlying space.
  • Entropy Regularization for Efficiency: Adding an entropy term to the OMT problem transforms it into a strictly convex optimization problem, enabling the use of efficient iterative algorithms like Sinkhorn iterations. This is crucial for practical applications.
  • Non-Coherent Fusion: The method is designed for scenarios where only aggregated data (covariance matrices) from separate sensor arrays are available, making it suitable for distributed systems with limited communication.
  • Robustness to Misalignment: The numerical results demonstrate that the proposed method is significantly less affected by array alignment errors compared to traditional techniques, showing a much slower degradation in performance as misalignment increases.

Key Results

The paper demonstrates the proposed method's robustness through numerical simulations in a 2D localization scenario. Two uncorrelated sources were simulated, impinging on two sensor arrays with varying degrees of misalignment.

  • Improved Robustness: The proposed method showed significantly higher robustness to misalignment compared to non-coherent MUSIC, MVDR, LS, and SPICE estimators. While the proposed method might have a slightly larger deviation from the ground truth in the perfectly aligned case, its performance degrades much more gracefully as the misalignment angle increases.
  • Qualitative Comparison: Figure 1 and 2 show that the proposed estimator produces a cleaner spectral estimate with less spurious peaks compared to the non-coherent MUSIC estimator when misalignment is present.

Limitations and Future Work

  • Computational Complexity: While entropy regularization improves efficiency, the computational cost can still be significant for very high-dimensional problems or a large number of sensor arrays. The paper mentions that the method is applicable to higher-dimensional problems, but practical limits may exist.
  • Noise Sensitivity: The paper uses a specific noise model (uncorrelated Gaussian noise). The performance with different noise characteristics or correlated noise might warrant further investigation.
  • Future Work: The authors suggest that the framework can be extended to include range estimation or handle higher spatial dimensions, indicating potential for more comprehensive localization solutions. Further exploration of the computational trade-offs and performance under various noise conditions would also be valuable.

Summary updated

Abstract

This work presents a method for information fusion in source localization applications. The method utilizes the concept of optimal mass transport in order to construct estimates of the spatial spectrum using a convex barycenter formulation. We introduce an entropy regularization term to the convex objective, which allows for low-complexity iterations of the solution algorithm and thus makes the proposed method applicable also to higher-dimensional problems. We illustrate the proposed method's inherent robustness to misalignment and miscalibration of the sensor arrays using numerical examples of localization in two dimensions.
BibTeX
@inproceedings{elvander2019non,
  author = {Filip Elvander and Isabel Haasler and Andreas Jakobsson and Johan Karlsson},
  title = {{Non-coherent Sensor Fusion via Entropy Regularized Optimal Mass Transport}},
  booktitle = {IEEE International Conference on Acoustics, Speech and Signal Processing},
  year = {2019},
  url = {https://openreview.net/forum?id=JOy9emtlRS},
}
Figure