⚡ electrical engineering

Efficient 2.5-D FEM-Based Scattering Analysis of the Human Body for RF Sensing

This paper proposes a fast 2.5-D Finite Element Method for efficiently simulating human body scattering under RF excitation to generate large-scale, labeled datasets for training Device-Free Localization and RF sensing systems, thereby overcoming the limitations of costly real-world data collection.

Original authors: Haoqing Wen, Michele D'Amico, Matteo Oldoni, Federica Fieramosca, Vittorio Rampa, Stefano Savazzi, Qi Wu, Gian Guido Gentili

Published 2026-03-18

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Haoqing Wen, Michele D'Amico, Matteo Oldoni, Federica Fieramosca, Vittorio Rampa, Stefano Savazzi, Qi Wu, Gian Guido Gentili

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to find a person hiding in a dark room using only a flashlight and a mirror. You can't see the person directly, but you can see how the light bounces off them and changes the pattern on the walls. This is essentially what Device-Free Localization (DFL) does, but instead of visible light, it uses invisible radio waves (like Wi-Fi) to "see" people without them wearing any special devices.

The problem? To teach a computer to recognize these invisible patterns, you need a massive library of examples. Usually, you'd have to hire hundreds of people, put them in a room, and measure the radio waves thousands of times. That's expensive, slow, and exhausting.

This paper introduces a super-fast, virtual "simulator" that creates these examples for you, so you don't have to do all the physical measuring.

Here is a breakdown of how they did it, using some everyday analogies:

1. The Problem: The "Full 3D Movie" is Too Heavy

To simulate how radio waves bounce off a human body, scientists usually use Full 3D Simulations. Think of this like trying to render a high-definition 3D movie of a person standing in a room. It looks perfect, but it takes a supercomputer days to calculate just one frame. If you need thousands of frames to train an AI, you'd be waiting forever.

2. The Solution: The "Spinning Top" Trick (2.5-D FEM)

The authors realized that a human body, roughly speaking, looks the same if you spin it around (like a spinning top or a cylinder). They used a clever mathematical shortcut called 2.5-D FEM (Finite Element Method).

The Analogy: Imagine you want to know the shape of a vase. Instead of modeling every single curve in 3D space, you just draw the profile of the vase on a piece of paper and tell the computer, "Spin this line around."
The Result: This turns a massive, heavy 3D calculation into a much lighter 2D calculation. It's like switching from rendering a full 3D movie to just calculating the shadow of a spinning object. It's 100 times faster but still accurate enough to be useful.

3. The "Ghost" and the "Mirror"

In their simulation, they place a "virtual human" (modeled as a Body of Revolution) in a room with radio antennas.

The Dipole: They use a simple radio antenna (a dipole) as the light source.
The Ground: Since radio waves bounce off the floor, they had to account for that. They used a "mirror image" trick. Imagine the floor is a mirror; the computer pretends there is a "ghost person" underneath the floor reflecting the waves. This helps the simulation match real-world interference patterns without needing to know the exact chemical makeup of the floorboards.

4. The "Shaking Hand" (Micro-Movements)

One of the biggest challenges in RF sensing is that people aren't statues. They breathe, shift their weight, and fidget. These tiny movements change the radio signals slightly.

The Challenge: In real life, these tiny changes are messy and hard to label.
The Simulation: The authors used their fast simulator to generate thousands of "what-if" scenarios. They asked, "What happens to the signal if the person moves their hand 2 centimeters left? What if they breathe?"
The Discovery: They found that while the average signal strength doesn't change much with tiny movements, the difference in signal (the "ripple") is very sensitive. This is like noticing that a calm lake looks the same from far away, but if you zoom in, you can see the tiny ripples caused by a fish jumping.

5. Why This Matters: The "Training Gym"

The ultimate goal is to build AI that can track people in hospitals, smart homes, or disaster zones without them wearing sensors.

Before: You had to build a gym, hire actors, and run thousands of drills to train the AI.
Now: You can use this 2.5-D Simulator as a "virtual gym." You can generate millions of labeled training examples in minutes. The AI learns the physics of how radio waves bounce off humans, and then you just need a little bit of real-world data to fine-tune it.

Summary

This paper is about building a fast, virtual wind tunnel for radio waves. Instead of building a real wind tunnel and blowing air on real people to see how the wind moves, they built a computer model that spins a simplified human shape to calculate the wind patterns instantly.

This allows engineers to create huge datasets to train AI systems, making "invisible" radio sensing faster, cheaper, and more accurate for real-world applications like health monitoring and smart home security.

1. Problem Statement

Device-Free Localization (DFL) and Radio-Frequency (RF) sensing rely heavily on machine learning models trained with large-scale datasets. However, acquiring these datasets through physical measurements is:

Costly and Time-Consuming: Requires extensive campaigns across multiple individuals, environments, and device configurations.
Labor-Intensive: Difficult to cover diverse body poses, micro-movements, and environmental conditions.
Simulation Limitations: Existing numerical methods face a trade-off between accuracy and efficiency:
- Full-wave 3-D methods (FEM, FDTD, MoM): High accuracy but computationally prohibitive for large-scale database generation.
- Analytical/Semi-analytical models (Spheres, Cylinders): High efficiency but lack realism in representing complex human scattering.
- High-frequency approximations: Efficient for large-scale propagation but inaccurate for near-field scattering and asymmetric excitations.

The core challenge is to develop a simulation method that is computationally efficient enough to generate large-scale labeled datasets while maintaining high numerical accuracy to capture realistic human scattering effects, including micro-movements.

2. Methodology

The authors propose a fast 2.5-D Finite Element Method (FEM) tailored for a Body of Revolution (BoR) human model excited by a z-directed dipole.

Key Technical Components:

BoR Approximation: The human body is modeled as a Body of Revolution. This exploits rotational symmetry to decouple the electromagnetic field into circular harmonics ( $m$ ) in the azimuthal direction ( $\phi$ ).
2.5-D Formulation:
- The problem is solved in a 2D cross-sectional plane ( $\rho, z$ ) rather than a full 3D volume.
- The scattered electric field is expanded as a sum of circular harmonics: $E(\rho, \phi, z) = \sum E^{(m)}(\rho, z) \cdot \text{trig}(\phi)$ .
- This transforms the 3D vector problem into a series of independent 2D problems for each harmonic $m$ .
Galerkin's Method & Mixed Basis:
- The in-plane field components are expanded using edge-based basis functions.
- The out-of-plane component is expanded using nodal basis functions.
- This results in a block-diagonal system of equations, where each block corresponds to a specific harmonic. This structure allows for massive parallelization or sequential solving of small matrices, drastically reducing memory and time.
Excitation and Boundary Conditions:
- Source: A z-directed Hertzian dipole is placed outside the FEM domain.
- Ground Plane: A Perfectly Matched Layer (PML) simulates free space, while an image source method with a synthetic reflection coefficient ( $\Gamma$ ) approximates the ground reflection effects.
- Equivalent Sources: To compute fields at distant receivers without extending the FEM mesh, equivalent electric and magnetic currents are deployed on a cylindrical surface surrounding the body. These are used to calculate the field in the far-field region via surface integrals.
Material Properties: Human tissue is modeled using the Cole-Cole model (6-term, 14 parameters) to define complex relative permittivity ( $\epsilon_r$ ) at the operating frequency (2.43 GHz).

3. Key Contributions

Efficient 2.5-D FEM Formulation: Development of a specific 2.5-D FEM framework for dipole-excited BoR human models. It reduces computational complexity by a factor of at least $M^2$ (where $M$ is the number of harmonics, typically $\approx 10$ ) compared to full 3-D FEM, while maintaining high accuracy.
Validation of Indoor RSSI Prediction: Successful validation of the method against both full-wave commercial simulations (Feko) and real-world indoor DFL experiments. The method accurately predicts spatial Received Signal Strength Indicator (RSSI) distributions.
Synthetic Dataset Generation Tool: Demonstration that the method can generate synthetic RF data that statistically matches real measurements. Specifically, it models human micro-movements (small displacements and breathing) to analyze their impact on RSSI statistics, providing a tool for creating labeled training datasets for machine learning.

4. Results

Numerical Accuracy:
- Comparison with Feko (Method of Moments) showed excellent agreement in normalized directivity patterns.
- Speed: For a typical case, the 2.5-D FEM took 4.6 seconds (vs. 7 seconds for Feko). For larger, more complex indoor scenarios, the 2.5-D FEM took 60 seconds, whereas Feko required several hours, making the proposed method orders of magnitude faster for large-scale generation.
Experimental Validation:
- Experiments were conducted in an indoor room with 20 antennas (Tx/Rx) and a human subject at 5 positions.
- RSSI Correlation: Calibrated simulated RSSI values showed strong agreement with measured data across different transmitter-receiver links. The simulation correctly captured spatial trends, signal decay, and interference patterns (diffraction and shadowing).
Micro-Movement Analysis:
- The study analyzed the effect of small human displacements (within a $0.4m \times 0.4m$ area).
- $\Delta$ RSSI Sensitivity: While absolute RSSI showed limited sensitivity to micro-movements, the differential RSSI ( $\Delta$ RSSI) showed significant variation. The simulation successfully reproduced the Probability Density Functions (PDFs) of RSSI fluctuations observed in experiments, capturing both mean trends and statistical spreads.

5. Significance and Impact

Bridging the Data Gap: The method provides a practical, efficient solution to the "data hunger" of ML-based DFL systems. It enables the generation of large-scale, labeled, physics-based datasets without the prohibitive cost of physical measurement campaigns.
Calibration and Tuning: It serves as a reliable tool for calibrating tomographic reconstruction methods and tuning RF sensing algorithms by providing ground-truth data on how human geometry and micro-movements affect signal propagation.
Scalability: By reducing the computational burden from 3D to 2.5D, it makes high-fidelity electromagnetic simulation feasible for complex indoor environments and multi-target scenarios (future work).
Realism: Unlike simplified analytical models, this approach retains the geometric complexity of the human body, ensuring that the generated data is representative of real-world scattering phenomena.

In conclusion, this paper presents a robust bridge between electromagnetic theory and data-driven RF sensing, offering a computationally efficient tool that preserves the physical fidelity necessary for training next-generation localization and sensing systems.