Spectrum Shortage for Radio Sensing? Leveraging Ambient 5G Signals for Human Activity Detection

This paper introduces Ambient Radio Sensing (ARS), a novel ISAC approach that repurposes ambient 5G signals for human activity detection via a passive self-mixing hardware architecture and a cross-modal learning framework, effectively overcoming spectrum scarcity while preserving privacy.

The Gist

Imagine you are trying to watch a movie in a crowded theater, but the projector is broken, and the screen is covered in fog. You can't see the actors, but you can hear their footsteps and feel the vibrations of their movements. Now, imagine if you could use those faint vibrations to reconstruct a perfect, high-definition video of the actors, even though you never actually saw them.

That is essentially what this paper, "Ambient Radio Sensing (ARS)," is about.

Here is the story of how they solved a major problem using a clever mix of hardware and software.

The Problem: The "Spectrum Traffic Jam"

Imagine the airwaves (the spectrum) as a busy highway.

• Traditional Radar is like a police car with its own siren and lights. It sends out a signal, waits for it to bounce back, and figures out where things are. But to do this, it needs its own dedicated lane on the highway.
• The Problem: The highway below 10 GHz (where most radio waves live) is already packed. It's full of 5G, Wi-Fi, TV, and emergency services. There is no room left for new radar lanes. If you try to build a new radar, you'll cause a traffic jam (interference) or get a ticket from the regulators.

The Solution: The "Echo Hunter" (ARS)

Instead of building a new radar with its own siren, the researchers built a device called ARS that acts like a passive echo hunter.

Think of ARS as a smart microphone that doesn't listen to voices, but listens to the echoes of other people's voices.

1. No Siren Needed: ARS doesn't generate its own radio waves. Instead, it "eavesdrops" on the 5G signals that are already flying around your house or office (like a radio station playing music in the background).
2. The "Flashlight" Trick: The ARS device catches these existing 5G signals, turns up the volume (amplifies them), and shines them like a flashlight onto the room.
3. The Echo: When these amplified signals hit a person walking by, they bounce back. ARS catches these bounces.
4. The Magic: Because the device knows exactly what the original signal looked like, it can compare the "bounced" signal to the "original" signal. The tiny differences tell it exactly how the person moved, how fast they were going, and even what their body shape looks like.

Why is this cool? It's like using the sunlight already in the room to take a photo, rather than needing to bring your own flashbulb. It doesn't interfere with the 5G network; in fact, it might even help boost the signal for the phone in your pocket!

The Hardware: The "Self-Mixing" Kitchen

The device uses a clever piece of hardware called a Self-Mixing RF architecture.

• Analogy: Imagine you are in a kitchen. You have a pot of soup (the 5G signal). You take a spoonful, taste it (the reference), and then take a spoonful of the soup after it has been stirred by a spoon (the reflection off a person).
• By mixing these two spoonfuls together, you can instantly tell exactly how the spoon moved the soup.
• In the real world, this "mixing" happens at the speed of light, turning the complex radio waves into a simple, readable signal that a computer can understand.

The Software: The "Vision Translator"

Here is the tricky part: Radio waves are messy. They are like a blurry, static-filled TV screen. Turning that blur into a clear picture of a human skeleton is hard.

To solve this, the researchers used a technique called Cross-Modal Learning.

• The Analogy: Imagine teaching a blind person to recognize shapes. You can't just show them a picture. Instead, you hold a 3D model of a cat in their hands while showing them a picture of a cat to a sighted person. You tell the sighted person, "This is what a cat looks like," and you let them "teach" the blind person what a cat feels like.
• In the Lab: They built a system where the ARS device and a regular video camera worked side-by-side. The camera (the "sighted" teacher) recorded the person's exact skeleton and body shape. The computer used this perfect video data to train the radio device (the "blind" student).
• The Result: Once trained, the radio device no longer needs the camera. It can look at the messy radio echoes and "dream" up a perfect skeleton and body mask, just like it learned from the video.

What Did They Prove?

They built a prototype and tested it in real rooms with real 5G signals.

• The Test: They asked people to walk around, wave their arms, and move in complex ways.
• The Outcome: The ARS device successfully drew the person's skeleton (like a stick figure) and traced their body outline (a "body mask") with high accuracy.
• The Privacy Win: Because it uses radio waves, it can see through walls and doesn't capture faces or details. It's perfect for nursing homes or hospitals where you want to monitor an elderly person's movement without invading their privacy with a camera.

The Bottom Line

This paper introduces a way to turn our crowded radio spectrum into a superpower. Instead of fighting for empty space to build new radars, we can use the 5G signals that are already everywhere to "see" people, track their movements, and monitor their health—all without cameras, without new spectrum licenses, and without interfering with our phone calls.

It's like turning the entire world's Wi-Fi and 5G network into a giant, invisible, privacy-preserving motion sensor.

Technical Summary

1. Problem Statement

Radio sensing in the sub-10 GHz spectrum offers significant advantages over vision-based systems (e.g., cameras, LiDAR), including the ability to operate in darkness, penetrate obstacles (walls, fog), and preserve user privacy. However, the widespread deployment of radio sensing is hindered by spectrum scarcity. The sub-10 GHz band is heavily allocated to critical services like 5G, Wi-Fi, and broadcasting. Introducing new dedicated sensing radars in this range faces strict regulatory hurdles and interference issues. While millimeter-wave (mmWave) radars offer high resolution, they suffer from poor obstacle penetration and short range.

Core Challenge: How to perform high-fidelity radio sensing (specifically human activity detection) in the sub-10 GHz band without requiring dedicated spectrum or interfering with existing communication systems.

2. Methodology: Ambient Radio Sensing (ARS)

The authors propose Ambient Radio Sensing (ARS), an Integrated Sensing and Communications (ISAC) approach that repurposes existing over-the-air signals (e.g., 5G, Wi-Fi) for sensing without disrupting their primary communication functions.

A. Hardware Architecture

The ARS device operates as a standalone unit using an Amplify-and-Forward and Self-Mixing architecture:

• Signal Reception & Illumination: An omnidirectional dipole antenna (RX0) receives ambient RF signals from a base station. These signals are amplified and re-transmitted via a directional patch antenna (TX) to illuminate the environment.
• Self-Mixing Circuit: Multiple receiving patch antennas (RX1–RXM) capture signals reflected from objects. Crucially, the system mixes these reflected signals with a copy of the amplified original signal (from the TX path) in the analog domain.
• Frequency Selectivity: The patch antennas are optimized to act as band-pass filters, suppressing out-of-band interference and ensuring the system only processes the specific communication band of interest.
• Advantage: By amplifying and forwarding the signal, ARS potentially boosts the communication signal strength for the network while generating a baseband signal suitable for feature extraction without generating new interference.

B. Signal Processing & Mathematical Foundation

• Doppler Extraction: The authors derive that the phase of the generated baseband signal exhibits a linear relationship with object displacement ($d_l$). Even with the limited bandwidth and OFDM structure of 5G/Wi-Fi, the phase rotation allows for the estimation of instantaneous velocity ($v_l$).
• Signal Sanitization: To handle noise, bias, and outliers inherent in ambient signals, a multi-stage pipeline is used:
  1. Filtering: Butterworth low-pass filter to remove high-frequency noise.
  2. Bias Correction: K-means clustering to identify and remove static signal offsets.
  3. Outlier Removal: Local Outlier Factor (LOF) algorithm to eliminate sparse, anomalous points caused by interference.
  4. Projection: K-means clustering reduces the constellation diagram to a single representative point per time frame.
• Differential Beamforming: To generate spatial heatmaps, the system uses a 2D antenna array. By computing the difference between signal frames over a time gap ($T_\Delta$), static components are suppressed, and moving objects are highlighted. This creates a heatmap $X(t)$ representing the spatial distribution of motion.

C. Cross-Modal Learning Framework

To overcome the sparsity and noise of radio features for fine-grained tasks (skeleton estimation and body segmentation), the authors propose a Cross-Modal Supervised Learning approach:

• Training Phase: The ARS system is co-located with a camera. A vision model (e.g., HRNet, Mask R-CNN) extracts ground-truth labels (skeletons, masks) from synchronized video.
• Model Architecture: A Deep Neural Network (DNN) processes the radio heatmaps.
  • Encoder: Uses a Transformer architecture with sequential patch projection to capture long-range spatio-temporal dependencies in the radio data.
  • Decoder: Uses deconvolution layers to upsample features for dense prediction (segmentation masks and keypoint heatmaps).
• Inference Phase: Once trained, the camera is discarded. The model relies solely on ambient radio features to perform privacy-preserving activity detection.

3. Key Contributions

1. Novel ISAC Paradigm: Introduces ARS, a method to solve spectrum scarcity by passively leveraging ubiquitous ambient signals (5G/Wi-Fi) for sensing without interfering with communication.
2. Joint Hardware-Algorithm Design: Proposes a self-mixing RF architecture that enables robust Doppler and angular feature extraction from ambient OFDM waveforms, validated by mathematical modeling of phase-distance relationships.
3. Cross-Modal Learning: Develops a framework that distills high-fidelity visual knowledge into radio models, enabling fine-grained human activity recognition (skeletons and masks) from sparse radio data.
4. Prototype & Validation: Built a functional prototype and validated it in real-world scenarios, demonstrating state-of-the-art performance compared to existing WiFi-based sensing methods.

4. Experimental Results

The authors evaluated ARS using a private 5G testbed (2.35 GHz, 40 MHz bandwidth) and a dataset of 206k+ training samples across diverse indoor/outdoor scenes.

• Performance Metrics:
  • Mask Segmentation: ARS achieved an Average Precision (AP) of 0.5097, outperforming baselines like SiWiS (0.4805) and Person-in-WiFi (0.3800). Notably, at high precision thresholds (AP@0.80), ARS showed a 150% improvement over SiWiS.
  • Keypoint Estimation: ARS achieved an AP of 0.3997 and AR of 0.5071, significantly outperforming SiWiS (AP: 0.3469).
• Robustness: The system maintained high fidelity even under stringent criteria, proving the self-mixing architecture captures fine-grained spatial features lost in traditional Channel State Information (CSI) methods.
• Physical Analysis:
  • Distance: Performance degrades as distance increases (2m to 6m) due to SNR reduction, but torso keypoints remain robust longer than extremities (wrists/ankles).
  • Duration: A 4.0-second signal window was found optimal for capturing complete kinematic cycles.
• Qualitative Results: Visualizations showed ARS-generated skeletons and masks closely matching ground-truth video, even in multipath-rich environments.

5. Significance

This work represents a paradigm shift in radio sensing by moving away from dedicated, spectrum-hogging radars toward spectrum-exempt, ambient sensing.

• Scalability: By utilizing existing 5G/Wi-Fi infrastructure, ARS removes the regulatory and cost barriers to deploying large-scale sensing networks.
• Privacy: It offers a privacy-preserving alternative to cameras for sensitive environments (hospitals, nursing homes) while retaining the ability to detect complex human activities.
• Future of ISAC: It demonstrates that communication signals can be effectively "recycled" for high-precision sensing, paving the way for future 6G networks where sensing and communication are deeply integrated.