Tensor Sensing for Rf Tomographic Imaging

Tao Deng, Feng Qian, Xiao-Yang Liu, Manyuan Zhang, Anwar Walid

TL;DR

This paper introduces a novel tensor sensing approach for RF tomographic imaging, enabling more accurate and efficient reconstruction of 3D physical spaces. By leveraging a transform-based tensor model and a fast alternating minimization algorithm (Alt-Min), the method significantly outperforms prior tensor-based compressed sensing techniques in both recovery error and convergence speed, particularly for large-scale 3D environments.

What Problem Does This Paper Address?

Radio-frequency (RF) tomographic imaging aims to infer the physical layout of objects and spaces by analyzing how RF signals propagate through them. Traditional methods often treat the data as vectors, which fails to capture the inherent geometric structures of 3D spaces. This limitation leads to lower accuracy in reconstructing complex environments. While some tensor-based methods exist, they suffer from high computational complexity due to operations like tensor singular value decomposition (t-SVD), making them impractical for large datasets. This paper addresses the need for a more efficient and accurate RF tomographic imaging technique that can effectively exploit the 3D geometric structure of the target space.

Methods

The core of this paper is a novel tensor sensing approach that formulates the RF tomographic imaging problem within a transform domain. Key methodological contributions include:

  • Transform-Based Tensor Model: The authors adopt a transform-based tensor model that leverages real-valued fast transforms (like FFT and DCT) to exploit the geometric structures of 3D sensory data. This model is more suitable for received signal strength (RSS) data and allows for diverse sampling strategies.
  • Tensor Sensing Formulation: The problem is framed as inferring a low L-rank tensor (representing the loss field) from a set of linear measurements.
  • Alternating Minimization (Alt-Min) Algorithm: A fast alternating minimization algorithm is proposed. This algorithm decomposes the loss field tensor into the product of two smaller tensors and iteratively estimates these tensors. This factorization significantly reduces computational complexity compared to methods requiring full tensor decompositions.
  • Algorithm Optimization: The implementation of Alt-Min is optimized to reduce memory and computation requirements, making it suitable for larger 3D spaces. This optimization involves using circulant algebra and transforming the problem into a series of standard least squares minimization problems.

Datasets

The authors evaluate their proposed Alt-Min approach using the IKEA 3D datasets. Specifically, they use 3D models of an IKEA chair and an IKEA table to generate ground truth tensors of size 60x60x15. These models are placed within a simulated 3D space to mimic real-world scenarios. The paper does not introduce a new benchmark dataset but rather uses existing 3D object models to create realistic ground truth data for evaluation.

Key Insights and Ideas

  • Geometric Structure is Key: The paper strongly emphasizes that exploiting the geometric structure of the 3D space is crucial for accurate RF tomographic imaging. Vector-based methods, by ignoring this structure, are inherently limited.
  • Low L-Rank Property: The transform-based tensor model exhibits a low L-rank property, meaning that most of the tensor's energy is captured by a small number of singular values. This is a significant advantage over traditional matrix SVD or CP-decomposition, which require many more components to represent the same energy.
  • Efficiency through Factorization: Decomposing the problem into estimating two smaller tensors via alternating minimization is a key to achieving both speed and accuracy. This avoids the high computational cost of full tensor decompositions.
  • Practical Implications for Large Spaces: The optimization techniques employed make the method practical for inferring large three-dimensional physical spaces, which is a significant step forward for RF tomography applications.

Key Results

  • Improved Recovery Error: The Alt-Min approach, both with FFT and DCT, demonstrates significantly lower recovery errors compared to prior tensor-based compressed sensing methods. For an 80% sampling rate, the RSE of Alt-Min is around 10⁻¹³ to 10⁻¹², while the prior method's RSE is around 10⁻³.
  • Faster Convergence: The Alt-Min algorithm converges much faster. For a target RSE of 10⁻¹⁰, approximately 27 iterations are needed for both Alt-Min variants, whereas the prior method's convergence is much slower.
  • Successful Outline Reconstruction: Qualitative results show that the Alt-Min with DCT successfully recovers the outlines of IKEA chair and table models, demonstrating its ability to capture spatial information.
  • Scalability: The optimized algorithm shows a significant reduction in space complexity, making it more feasible for inferring larger 3D physical spaces.

Limitations and Future Work

  • Focus on Outlines: While the paper demonstrates successful recovery of object outlines, the focus is primarily on this aspect rather than the full texture and color information of the objects.
  • Sampling Rate Sensitivity: The paper notes that for very low sampling rates (20-25%), the prior tensor-based compressed sensing method performs better. This suggests that the performance of Alt-Min might be more sensitive to extremely sparse sampling.
  • Future Work: The paper implicitly suggests future work in exploring different transform domains, further optimizing the algorithm for even larger scale problems, and potentially extending it to incorporate more detailed object properties beyond outlines. The authors also mention the potential for diverse sampling strategies, hinting at future research in adaptive or optimized sampling schemes.

Summary updated

Abstract

Radio-frequency (RF) tomographic imaging is a promising technique for inferring multi-dimensional physical space by processing RF signals traversed across a region of interest. However, conventional RF tomography schemes are generally based on vector compressed sensing, which ignores the geometric structures of the target spaces and leads to low recovery precision. The recently proposed transform-based tensor model is more appropriate for sensory data processing, as it helps exploit the geometric structures of the three-dimensional target and improve the recovery precision. In this paper, we propose a novel tensor sensing approach that achieves highly accurate estimation for real-world three-dimensional spaces. First, we use the transform-based tensor model to formulate a tensor sensing problem, and propose a fast alternating minimization algorithm called Alt-Min. Secondly, we drive an algorithm which is optimized to reduce memory and computation requirements. Finally, we present evaluation of our Alt-Min approach using IKEA 3D data and demonstrate significant improvement in recovery error and convergence speed compared to prior tensor-based compressed sensing.
BibTeX
@inproceedings{deng2018tensor,
  author = {Tao Deng and Feng Qian and Xiao-Yang Liu and Manyuan Zhang and Anwar Walid},
  title = {{Tensor Sensing for Rf Tomographic Imaging}},
  booktitle = {ICME},
  year = {2018},
  url = {https://openreview.net/forum?id=b7nVxjSKOj0},
}
Figure