MIRO: Multi-radar Identity and Ranging for Occupational Safety

MIRO is a privacy-preserving framework that combines distributed particulate matter sensors with a multi-radar mmWave re-identification system, utilizing GAN-based view adaptation to track workers and estimate their specific exposure to airborne pollutants in industrial environments without relying on visual data.

The Gist

Imagine you are working in a stone-cutting yard. It's loud, it's dusty, and the air is thick with tiny, invisible particles that can hurt your lungs over time. The big question for safety managers is: Who is breathing in the most dust, and why?

Traditionally, to answer this, you'd have two bad options:

1. Wear a mask with a computer in it: This is uncomfortable, gets dirty, and workers hate wearing it all day.
2. Put up security cameras: This feels like Big Brother watching you. Workers get nervous about their privacy, and it's often against the rules.

Enter MIRO (Multi-radar Identity and Ranging for Occupational Safety). Think of MIRO as a team of invisible, privacy-friendly "ghost eyes" that can see people and dust without ever taking a photo or touching a worker.

Here is how it works, broken down into simple concepts:

1. The "Ghost Eyes" (The Radars)

Instead of cameras that see your face and clothes, MIRO uses mmWave radars (the same kind of technology used in self-driving cars, but smaller).

• The Magic: These radars bounce radio waves off people. They can't see your face or what you are wearing, so your privacy is safe.
• The Superpower: They can see through dust, smoke, and darkness. While a camera would be blinded by a cloud of stone dust, the radar sees right through it.
• The Problem: A single radar is like a flashlight; it only sees a small slice of the room. If you have a big factory, you need many flashlights (radars) to cover the whole area.

2. The "Identity Crisis" (The Multi-Radar Puzzle)

Here is the tricky part: When you have multiple radars, they all see the same person, but they give them different names.

• Radar A sees a worker and calls him "User 11."
• Radar B, looking from a different angle, sees the same guy and calls him "User 24."

If the system doesn't know that "User 11" and "User 24" are the same person, it can't track how much dust that specific person breathed in. It's like having two different security guards who don't talk to each other; one thinks you are "John," and the other thinks you are "Mike," so they can't keep a single log of your day.

3. The "Translator" (The AI Brain)

This is where the paper's biggest innovation comes in. The researchers built a special AI translator (using something called a GAN, or Generative Adversarial Network).

• The Analogy: Imagine you are looking at a person through a window. If you look from the front, you see their face. If you look from the side, you see their profile. The picture looks totally different, but it's the same person.
• The Radar Challenge: For radars, looking from the side changes the "sound" of the person's movement (their Doppler signature). A grinding motion looks different from the side than from the front.
• The Solution: The AI acts like a universal translator. It takes the "side view" radar data and mathematically "rotates" it in its mind to look like the "front view" data. Once the data looks the same, the system can easily say, "Ah, User 11 and User 24 are definitely the same person!"

4. The "Dust Map" (Connecting People to Pollution)

Once the system knows exactly where every worker is (even if they move between different radar zones), it combines this with dust sensors scattered around the yard.

• The Result: The system creates a live, moving map. It doesn't just say, "There is dust in the factory." It says, "Worker Sarah, who is currently grinding stone near the back wall, is breathing in 500 units of dust per minute, while Worker Mike, who is standing near the door, is only breathing in 50."

Why This Matters

• Privacy First: No cameras, no facial recognition. Just movement patterns.
• Comfort: Workers don't have to wear heavy, sweaty sensors.
• Accuracy: It solves the "who is who" problem in big, dusty, multi-angle environments where other systems fail.

The Bottom Line

MIRO is like a smart, invisible safety net. It uses radio waves to track workers' movements, a super-smart AI to make sure it knows who is who from every angle, and dust sensors to calculate exactly how much pollution each person is exposed to. This allows factory owners to protect their workers' health without making them feel watched or uncomfortable.

In short: It keeps the dust out of your lungs and your privacy in your pocket.

Technical Summary

Here is a detailed technical summary of the paper "MIRO: Multi-radar Identity and Ranging for Occupational Safety."

1. Problem Statement

Occupational safety in open industrial workspaces (e.g., stone-cutting yards, construction sites) is severely compromised by exposure to airborne particulate matter (PM). Accurate risk assessment requires joint, time-aligned measurement of worker activity and localized pollutant concentration. However, existing monitoring solutions face critical limitations:

• Wearable Sensors: Intrusive, require maintenance, and can bias worker behavior under heavy physical loads.
• Camera-based Tracking: Raises severe privacy and compliance concerns, leading to worker resistance.
• Single-Sensor mmWave Radar: While privacy-preserving and robust to dust, a single radar has limited Field of View (FoV). Monitoring large or irregular spaces requires multiple overlapping radars, which introduces cross-view identity ambiguity (the same worker is assigned different local IDs by different radars).
• Viewpoint Sensitivity: Micro-Doppler signatures (used for identification) are highly dependent on the viewing angle (azimuth), making direct comparison of data from different radars unreliable.

Goal: Develop a privacy-preserving, device-free system that tracks workers across multiple radar viewpoints to enable personalized PM exposure estimation without relying on visual cues.

2. Methodology

The MIRO framework integrates continuous PM sensing with a multi-radar millimeter-wave (mmWave) re-identification (re-ID) backbone. The system architecture consists of four main stages:

A. User Localization (TDSCAN)

Standard clustering algorithms (like DBSCAN) fail in multi-radar settings due to sparse data and temporal inconsistencies. MIRO introduces TDSCAN (Temporal Doppler Spatial Clustering):

• Doppler Filtering: Filters out static clutter by retaining only points with non-zero Doppler velocities (specifically within a range of 0.05 m/s to 1.0 m/s), isolating human motion.
• Temporal Consistency: Aggregates detections over a short temporal window (10 frames) and uses the Hungarian algorithm to associate centroids across frames, ensuring continuous tracks even with transient signal loss.

B. Activity Signature Extraction

For each localized user, the system extracts a Range-Doppler (RD) heatmap centered on the user's range bin. This isolates the micro-Doppler signatures caused by limb movements and tool interactions, filtering out background clutter and static machinery.

C. View-Adaptation via Generative AI

To solve the cross-view identity problem, MIRO employs a Pix2Pix-based Generative Adversarial Network (GAN):

• Architecture: A U-Net generator with skip connections and a PatchGAN discriminator.
• Training: Trained on paired RD data from two radars at different azimuthal angles (0° to 360°).
• Function: The network learns to translate an RD signature from one radar's viewpoint to another's. It specifically models azimuthal distortions (since most stone-working motions occur in the horizontal plane) while preserving intrinsic motion patterns.
• Loss Function: Uses Least-Squares GAN (LSGAN) for stability and an $L_1$ reconstruction loss to ensure structural fidelity.

D. Cross-Radar Association & Re-ID

• Correlation Matching: After view adaptation, the system computes the normalized correlation between the adapted signature of a user in Radar A and the observed signature in Radar B.
• Global Identity: Pairs with high correlation are linked using the Hungarian algorithm. These links are aggregated into a graph where connected components define a Global User ID (e.g., $P_1$), ensuring identity persistence across the entire network.

E. Personalized Exposure Estimation

Once global trajectories are established, the system maps worker positions to a spatially interpolated PM field generated by a distributed network of DALTON PM sensors. Personalized exposure ($E_k$) is calculated as the time-averaged PM concentration along each worker's trajectory.

3. Key Contributions

1. MIRO Framework: A novel multi-radar re-ID system that combines robust localization (TDSCAN) with an azimuth-aware Pix2Pix view-adaptation network.
2. View Adaptation Mechanism: The first application of GAN-based image-to-image translation to radar RD signatures, enabling the system to generalize across unseen workspaces and activity types without retraining.
3. Cross-Radar Association: A correlation-based engine that ensures global identity consistency across overlapping radar fields, overcoming the lack of visual cues or MAC addresses.
4. Real-World Validation: Comprehensive evaluation in three distinct, harsh environments (outdoor marble factory, semi-mechanized stone-cutting factory, and indoor construction site) demonstrating robustness against dust, multipath interference, and mechanical vibrations.

4. Results

The system was evaluated in both controlled laboratory settings and real-world field trials.

• Re-Identification Performance:
  • Achieved a 90.4% F1-score for cross-radar re-ID in lab tests.
  • Outperformed baseline methods (distance-based and raw correlation) by ~20%, proving the necessity of view adaptation.
  • Maintained high performance across varying user densities (2–4 users).
• View Adaptation Accuracy:
  • Achieved a mean Structural Similarity Index Measure (SSIM) of 0.70 and a PSNR of 28.02 dB using the Pix2Pix-LSGAN model, significantly outperforming CycleGAN and Pix2Pix-WGAN-GP.
  • Demonstrated strong generalization to unseen activities (clapping, waving) and unseen environments without fine-tuning.
• Localization Accuracy:
  • Mean Absolute Error (MAE) for user positioning was < 10 cm compared to Vicon motion capture ground truth.
• Exposure Analysis:
  • Successfully differentiated PM exposure levels based on activity (e.g., Grinding: ~600 $\mu g/m^3$ PM2.5 vs. Polishing: ~100 $\mu g/m^3$).
  • Revealed that "passive" workers near high-dust activities still face significant exposure, validating the need for trajectory-based monitoring.
• Latency:
  • End-to-end system latency is ~230 ms (excluding sensor aggregation), which is acceptable for occupational safety monitoring where activity changes occur over minutes.

5. Significance

MIRO addresses a critical gap in industrial safety by providing a privacy-preserving, maintenance-free, and robust solution for monitoring occupational health.

• Privacy: Unlike cameras, it does not capture visual identities, addressing worker resistance and regulatory hurdles.
• Robustness: Unlike wearables, it functions in high-dust, high-noise environments where optical sensors fail and wearables are uncomfortable.
• Precision: It moves beyond coarse ambient monitoring to provide worker-specific exposure estimates, enabling targeted safety interventions.
• Scalability: The view-adaptation approach allows the system to scale to large, multi-radar deployments in complex industrial settings without requiring retraining for every new site.

The paper concludes that mmWave sensing, when augmented with deep learning for view adaptation, is a viable and superior alternative for next-generation occupational safety monitoring in harsh industrial environments.