⚡ electrical engineering

Human Walking Sensing and Pose Estimation in the 6 GHz Band Using Amplitude and Phase CSI

This paper presents a deep learning-based pipeline that leverages both amplitude and phase Channel State Information from 6 GHz OFDM signals in a multistatic network to achieve reliable human pose estimation, demonstrating that DT-Pose offers the highest accuracy and that phase data serves best as a complementary feature to amplitude.

Original authors: Zhaorui Yin, Mattia Brambilla, Monica Nicoli

Published 2026-06-10

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Zhaorui Yin, Mattia Brambilla, Monica Nicoli

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 in a room with three friends holding walkie-talkies. You start walking around. Even though you aren't holding a camera or wearing a smartwatch, your friends can figure out exactly how your arms and legs are moving just by listening to how your body bounces their radio signals around the room.

This paper is about teaching computers to do exactly that, but with much more advanced technology. Here is the breakdown of how they did it, using simple analogies.

The Setup: The "Radio Echo" Room

The researchers set up a small, 3x3 meter room (about the size of a large closet). Inside, they placed three special radio devices (called USRPs) that act like a team of echolocation experts.

The Signal: Instead of sending a simple "beep," these devices send out a complex, high-speed stream of radio waves (called OFDM) in the 6 GHz band. Think of this as a super-detailed, invisible fog filling the room.
The Disturbance: When a human walks through this fog, their body blocks, bounces, and twists the waves. Just like a rock thrown into a pond changes the water's ripples, a walking person changes the radio waves.
The Data: The devices catch these changed waves. They measure two things:
1. Amplitude: How strong the signal is (like the loudness of an echo).
2. Phase: The exact timing of the wave's cycle (like the precise position of a wave's peak).

The Challenge: Turning "Noise" into a Skeleton

The raw data looks like a chaotic mess of numbers. The goal is to turn this mess into a stick-figure skeleton that shows where the elbows, knees, and shoulders are.

To do this, the researchers used Deep Learning (AI). They took four existing "brain" models (named DT-Pose, MetaFi++, HPE-Li, and VST-Pose) that were originally trained to read Wi-Fi signals and retrained them to read these new 6 GHz radio signals.

The "2-in-1" Trick:
Most previous systems only looked at the "loudness" (amplitude) of the signal. This team decided to feed the AI both the loudness and the timing (phase) at the same time. They treated the radio data like a 3D puzzle, feeding the AI a massive block of information (9 radio links × 1,920 frequency channels × 100 time snapshots) to solve.

The Results: What Did They Find?

They tested the system by having a person walk around the room and comparing the AI's guess to a real, high-tech motion-capture suit (the "Ground Truth").

It Works: The AI successfully reconstructed a human skeleton. It could tell where the joints were, even though the person was just walking.
The "Blurry" Reality: The AI is good at figuring out the shape of the pose (e.g., "the arm is bent"), but it's not perfect at knowing the exact location in the room. The skeleton might be drawn 50 cm (about 20 inches) away from where the person actually is, but the pose itself looks correct.
Amplitude vs. Phase:
- Amplitude (Loudness): This was the heavy lifter. Using just the signal strength gave very good results.
- Phase (Timing): This was the "secret sauce." On its own, the timing data was confusing and didn't work well. However, when added to the loudness data, it helped refine the picture.
- The Exception: One model (MetaFi++) was actually better using only the timing data, but for the other three models, the timing data was only helpful as a sidekick to the loudness data.

The Bottom Line

The paper proves that we can use standard radio waves (specifically in the 6 GHz band used by modern Wi-Fi and 5G) to "see" people without cameras.

Privacy: No cameras are needed, so there are no video recordings of people.
Accuracy: It can build a reliable "stick figure" of a walking person.
Best Approach: The most accurate method combines signal strength with timing data, though signal strength alone does most of the heavy lifting.

In short, the researchers showed that by listening to the "echoes" of radio waves bouncing off a walking person, a computer can learn to draw a surprisingly accurate stick-figure of that person's movements.

Technical Summary: Human Walking Sensing and Pose Estimation in the 6 GHz Band Using Amplitude and Phase CSI

Problem Statement
This paper addresses the challenge of device-free human pose estimation using wireless signals in the emerging 6 GHz license-exempt band (5.925–7.125 GHz). While prior research has demonstrated the viability of using Channel State Information (CSI) from 2.4 GHz and 5 GHz Wi-Fi for sensing, these approaches rely on specific physical-layer parameters (e.g., channelization, signal format) inherent to Wi-Fi standards. The 6 GHz band, utilized by both Wi-Fi 6E and unlicensed 5G NR (NR-U), introduces different resource structures and waveforms, necessitating a re-evaluation of CSI-to-pose algorithms. A specific limitation in existing CSI-based pose estimation is its sensitivity to continuous human motion, such as walking, which often increases estimation errors and complicates simultaneous pose recovery and location tracking. Furthermore, the utility of phase information in this specific frequency band and bandwidth context remains under-explored compared to amplitude data.

Methodology
The authors propose a processing pipeline to estimate the pose of a human walking within a multistatic wireless network operating at 6 GHz. The system utilizes an open-access dataset collected in a $3 \times 3$ m $^2$ indoor area using three Universal Software Radio Peripherals (USRPs) configured in a multistatic setup. The transmitters emit Orthogonal Frequency-Division Multiplexing (OFDM) signals with a 16 MHz bandwidth at carrier frequencies of 6.74, 6.80, and 6.86 GHz.

The core methodology involves the following steps:

CSI Extraction: The Channel Impulse Response (CIR) is estimated from received OFDM signals using a Least-Squares (LS) approach. This yields a CSI tensor containing both amplitude and phase information across 1920 usable subcarriers and 100 OFDM symbols per frame for 9 distinct radio links.
Preprocessing: The pipeline processes the amplitude and phase components separately before fusion. Amplitude data undergoes discrete wavelet filtering to suppress high-frequency noise, while phase data is unwrapped and denoised using PhaseFi.
Feature Fusion: A 2-in-1 Multilayer Perceptron (MLP) fuses the amplitude and phase features into a unified input tensor of dimension $9 \times 1920 \times 100$ .
Deep Learning Architectures: Four state-of-the-art deep learning models, originally designed for Wi-Fi, are adapted to handle the higher resolution of the 6 GHz OFDM CSI. These models are:
- MetaFi++: A CNN-Transformer architecture using ResNet for feature extraction and Multi-Head Self-Attention (MHSA) for integration.
- DT-Pose: Utilizes Masked Autoencoder (MAE) style pretraining and a topology-constrained decoder combining MAE with Graph Convolutional Networks (GCN).
- HPE-Li: Replaces explicit transformers with dual-domain attention implemented via convolution (DSKConv).
- VST-Pose: Employs a dual-stream spatiotemporal transformer with an explicit velocity branch.
Pose Reconstruction: The models output 3D coordinates for 8 anatomical keypoints (chest, elbows, knees). To achieve a standard human representation, the system geometrically extrapolates these to 13 keypoints (adding hands, feet, and head) based on limb extension and swing velocity.

Key Contributions

Adaptation to 6 GHz OFDM: The paper successfully adapts existing Wi-Fi-centric deep learning models to process OFDM CSI from the 6 GHz band, validating their applicability in non-Wi-Fi waveform scenarios.
Systematic Evaluation of Phase Information: A primary contribution is the comparative analysis of amplitude-only, phase-only, and joint amplitude-phase CSI representations. The study systematically tests whether phase information acts as a standalone feature or a complementary one in this specific frequency and bandwidth context.
Pipeline Validation: The authors design and validate a complete processing strategy, including noise suppression, feature fusion, and the modification of input layers in DL models to accommodate the specific subcarrier and symbol counts of the dataset.

Results
Performance was evaluated on a validation set using Procrustes-aligned Mean Per-Joint Position Error (PA-MPJPE) and Bone Length Loss (BLL).

Model Performance: Among the four architectures, DT-Pose achieved the best overall accuracy, with an amplitude-only PA-MPJPE of 6.4 cm (on the original 8 keypoints).
Amplitude vs. Phase: The results indicate that amplitude-only CSI yields performance comparable to joint amplitude-phase processing for most models. Phase information is generally more beneficial as a complementary feature rather than a standalone input.
Exception: MetaFi++ was the only model that demonstrated better results when using phase-only CSI compared to amplitude-only, suggesting model-specific sensitivities to phase data.
Absolute vs. Relative Accuracy: While the models successfully reconstructed plausible human skeletons (low BLL), the absolute position error remained significant (approximately 50 cm). The models primarily recover relative pose configurations rather than precise absolute locations within the room.

Significance
The paper claims that reliable human pose reconstruction can be achieved using 6 GHz OFDM CSI measurements, demonstrating that CSI-based sensing methods can be successfully extended from conventional Wi-Fi to emerging 6 GHz ISAC (Integrated Sensing and Communication) systems. The findings suggest that while amplitude carries the majority of discriminative content for pose estimation, phase measurements offer valuable complementary information, particularly when integrated into specific deep learning architectures. This work supports the feasibility of using shared-spectrum 6 GHz bands for non-intrusive, device-free sensing applications such as assisted living and industrial safety.

The Setup: The "Radio Echo" Room

The Challenge: Turning "Noise" into a Skeleton

The Results: What Did They Find?

The Bottom Line

More like this