Radar Autofocus Using Sparse Blind Deconvolution

Hassan Mansour, Dehong Liu, Petros T. Boufounos, Ulugbek S. KamilovICASSP 2018 · DOI · OpenReview

TL;DR

This paper tackles the challenge of accurately imaging a scene with radar when the exact positions of the antennas are unknown (position ambiguity). They propose a novel approach that reformulates this problem as a "sparse blind deconvolution" in the image domain, where antenna position errors are treated as spatial shifts. This method demonstrates superior performance compared to existing techniques, especially in noisy conditions.

What Problem Does This Paper Address?

High-resolution radar imaging, particularly with distributed or mobile antennas, faces a significant hurdle: antenna position ambiguity. Even with advanced navigation systems, the precise locations of antennas can be uncertain, leading to distortions in the radar image. Traditional "autofocus" methods often model these errors as global phase shifts in the frequency domain. However, this model breaks down in near-field imaging and imposes strict limitations on allowable antenna movements. The authors aim to overcome these limitations by developing a more general and robust autofocus technique.

Methods

The core innovation lies in reframing the radar autofocus problem. Instead of modeling position errors as frequency-domain phase shifts, the authors propose treating them as two-dimensional spatial shift operators in the image domain. This transforms the problem into a multichannel blind deconvolution task. In essence, the observed radar measurements are seen as a "true" radar image convolved with a unique spatial shift kernel for each antenna, representing its specific position error.

To solve this, they introduce a block coordinate descent framework. This algorithm iteratively refines two key components:

  1. The radar image (x): This is recovered using a fused Lasso penalty, which encourages sparsity (few reflective points) and piece-wise smoothness (objects have defined edges) in the image.
  2. The spatial convolution kernels (hm): These represent the antenna position errors and are also encouraged to be sparse (one-sparse property, meaning they are essentially pure shifts).

The optimization process alternates between updating the image and updating the kernels, leveraging techniques like FISTA (Fast Iterative Shrinkage-Thresholding Algorithm) and FPGD (Fast Projected Gradient Descent).

Datasets

The paper evaluates their method using:

  • Simulated data: They generated ten different simulation layouts with varying target configurations and antenna position perturbations, tested under different noise levels (PSNR from 4dB to 20dB). They also compared their method against a state-of-the-art iterative perturbation estimation scheme.
  • Experimental data: They used a real radar setup with a horn antenna and three cylindrical reflectors, collecting measurements from a distributed array.

Key Insights and Ideas

  • Image-domain convolution is more general: The key insight is that modeling antenna position errors as 2D spatial shifts in the image domain is more accurate and general than the traditional frequency-domain phase error model, especially for near-field scenarios.
  • Blind deconvolution as a unified framework: This reframing elegantly unifies the autofocus problem with blind deconvolution, allowing the use of powerful sparse recovery techniques.
  • Sparsity and smoothness are powerful priors: Leveraging the inherent sparsity and piece-wise smoothness of radar scenes, along with the one-sparse nature of shift kernels, is crucial for solving the otherwise ill-posed problem.
  • Robustness to noise: The proposed method demonstrates significantly better performance in the presence of noise compared to existing techniques.

Key Results

  • Superior image quality: Both simulated and experimental results show that the proposed method recovers focused radar images with fewer artifacts and better target localization compared to methods that ignore position errors or use traditional autofocus techniques.
  • Robustness to noise: Receiver Operating Characteristic (ROC) curves show that the proposed method maintains high detection performance across a wide range of noise levels, outperforming both measurement-domain blind deconvolution and fused Lasso recovery without autofocus.
  • Accurate kernel recovery: The recovered spatial shift kernels (Figure 13) are shown to be consistent with the true antenna perturbations, validating the model.

Limitations and Future Work

  • Computational complexity: While the block coordinate descent approach is efficient, the paper notes that the SparseLift formulation for blind deconvolution can become intractable for large datasets.
  • Convex formulation: The current optimization framework is nonconvex. The authors suggest that future work will focus on deriving a convex formulation of the image-domain blind deconvolution problem for guaranteed convergence.
  • Generalization: While the paper focuses on a specific signal model, exploring its applicability to more complex radar scenarios or different types of antenna perturbations could be a future direction.

Summary updated

Abstract

The radar autofocus problem arises in situations where radar measurements are acquired of a scene using antennas that suffer from position ambiguity. Current techniques model the antenna ambiguity as a global phase error affecting the received radar measurement at every antenna. However, the phase error signal model is only valid in the far field regime where the position error can be approximated by a one dimensional shift in the down-range direction. We propose in this paper an alternate formulation where the antenna position error is modeled using a two-dimensional shift operator in the image-domain. The radar autofocus problem then becomes a multichannel two-dimensional blind deconvolution problem where the static radar image is convolved with a two dimensional shift kernel for each antenna measurement. We develop an alternating minimization framework that leverages the sparsity and piece-wise smoothness of the radar scene, as well as the one-sparse property of the two dimensional shift kernels.
BibTeX
@inproceedings{mansour2018radar,
  author = {Hassan Mansour and Dehong Liu and Petros T. Boufounos and Ulugbek S. Kamilov},
  title = {{Radar Autofocus Using Sparse Blind Deconvolution}},
  booktitle = {IEEE International Conference on Acoustics, Speech and Signal Processing},
  year = {2018},
  doi = {10.1109/icassp.2018.8462402},
}
Figure