We are actively hiring multiple roles, please inquire here.

Crius.Co

Long-Range Heartbeat Monitor

Cover Image for Long-Range Heartbeat Monitor
Carlos Mulone
Carlos Mulone

Listening Beyond Sight

Imagine a sensor that can sit on a ridge, in an urban rooftop hide, or on a drone wing, and listen for the sub-millimeter vibrations of a human heartbeat from over five kilometers away. This system adapts its sensing parameters in real time—adjusting gain, phase alignment, filtering coefficients, and beam direction—to maintain a clean heartbeat track in the presence of wind gusts, engine rumble, and multipath reflections.

Over 1,000 hours of field trials have produced detailed environmental response curves for variables such as temperature (–20 °C to +60 °C), humidity (10 % to 90 % RH), and surface reflectivity (glass, metal, painted concrete). We log every sensor channel at full sampling rate and feed the data back into offline analysis pipelines, extracting insights that refine our calibration tables and AI hyperparameters. Those refinements are pushed to our edge modules as incremental firmware updates, ensuring each pod becomes more robust with every deployment.

“We treat every heartbeat detection as a data point in a living model of the environment—continually learning and adapting.”

This level of performance unlocks use cases across search and rescue (locating buried survivors through rubble), tactical overwatch (distinguishing friend from foe in the dark), and medical monitoring (vital signs assessment without contact).

The Physics of a Beat

Detecting a heartbeat at range leverages fundamental principles of wave physics in three domains:

  • Micro-Doppler Radar
    A chest-wall velocity of ±0.5 mm per beat induces a Doppler shift

    f_d = 2·v / λ ≈ (2 × 0.0005 m/s) / 0.03 m ≈ 33 Hz

    where v≈0.0005 m/s and λ=0.03 m (10 GHz), yielding f_d≈33 Hz. Our radar transceiver uses FMCW chirps with adjustable slope (up to 300 MHz/µs) and instantaneous bandwidth of 7.5 GHz to achieve range resolution under 2 cm and velocity resolution under 0.01 m/s.

  • Coherent Laser Vibrometry
    A coherent infrared beam at 1550 nm pointed at a reflective surface will undergo a phase shift proportional to the surface displacement. In plain terms:

λ is the laser wavelength (1.55 × 10⁻⁶ m)
Δx is the surface movement in meters
Δφ is the resulting phase shift in radians

Key system details:

  • Active Beam Stabilization: Piezo-driven tip/tilt mirrors keep the laser spot on target within 0.2 μrad, compensating for wind, platform motion, and atmospheric turbulence.
  • High-Sensitivity Detection: A single-photon avalanche photodiode (SPAD) array provides shot-noise-limited readout, ensuring we capture minute phase changes.
  • Displacement Sensitivity: Better than 0.1 μm at 2000 m range and under 0.5 μm out to 5000 m in clear-air conditions.
  • On-Board Diagnostics: Real-time monitoring of laser coherence, detector health, and alignment status to guarantee reliable operation in the field.

By converting sub-micron chest-wall vibrations into measurable phase shifts, this laser vibrometry channel lets us “see” heartbeats through walls, foliage, or vehicle armor when combined with our other sensing modalities.

  • Acoustic Propagation
    Heartbeat acoustics concentrate energy in the 20–200 Hz band. We characterized directional sound levels across substrates—open air (attenuation 0.05 dB/m at 100 Hz), wood, concrete, and soil. This informs our adaptive beamforming weights, which steer array nulls around dominant noise sources and maintain SNR gains above 20 dB over single-mic setups.

Combining these physics models yields a probabilistic sensor fusion framework that weights each modality by its real-time SNR estimate, rejecting spurious detections below confidence thresholds.

Sensor Subsystem Architecture

Our field-deployable pod houses three tightly integrated sensor modules:

  1. Laser Vibrometry Module

    • Single-frequency 1,550 nm DFB laser with coherence length >10 m
    • Piezo-actuated beam steering mirrors for jitter <0.1 µrad
    • InGaAs SPAD receiver with dual-quadrature demodulation
    • Real-time beam health diagnostics (power, mode-hop monitoring)

  2. Ultra-Wideband Radar Array

    • 4×4 patch antenna phased array (gain 18 dBi/element)
    • SDR front-end covers 3.1–10.6 GHz with 12-bit ADC at 5 GS/s
    • FPGA implements range-Doppler processing and clutter suppression
    • IMU-aided motion compensation for platform vibrations

  3. Acoustic Microphone Array

    • 16 MEMS capsules (sensitivity –26 dB FSPL, 20–200 Hz flat response)
    • FPGA-driven delay-and-sum beamformer with adaptive null steering
    • High-resolution 24-bit sigma-delta ADCs at 48 kHz sampling
    • Environmental noise classifier for dynamic gain control

All modules synchronize via a GPS-disciplined oscillator (jitter <10 ns) and share data over a 10 GbE backplane. A confidence-weighted voting algorithm cross-validates detections—heartbeats confirmed by at least two sensors are published as geolocated tracks with quality metrics.

Signal Processing Pipeline

The edge-AI pipeline executes five major stages in parallel, all within a 50 ms budget:

  1. Signal Conditioning

    • Temperature and gain calibration using onboard sensors
    • DC offset removal and analog filter compensation
    • Rolling noise-floor estimation with overlapping 50 ms windows

  2. Time-Frequency Decomposition

    • Multi-level Daubechies-4 discrete wavelet transform isolates 0.8–2 Hz components
    • Custom FIR filters reject out-of-band artifacts (wind, machinery)

  3. Neural Feature Extraction

    • 1D CNN with 8 residual blocks (kernel size = 5, channels = 32→128)
    • Quantized to 8 bit for TPU acceleration
    • Outputs 128-D embeddings capturing rise time, amplitude ratio, spectral skew

  4. Probabilistic Tracking

    • Joint Probabilistic Data Association Filter with ellipsoidal gating
    • Maintains N-track hypotheses per beat, prunes by scoring likelihoods
    • Updates track covariance and handles track initiation/termination

  5. Adaptive Thresholding

    • Reinforcement-learning agent adjusts detection thresholds based on reward (true positive rate vs. false positive penalty)
    • Converges within 5 ms to maintain target false-alarm rates under 1 per hour

Each pod logs raw and processed data to a circular buffer, enabling post-mission audit, algorithm validation, and continuous improvement.

Edge AI and System Integration

The compute core is an octa-core ARM Cortex-A72 (2.0 GHz) paired with an 8 TOPS TPU. Key integration details:

  • Operating System

    • Real-time Linux kernel with PREEMPT_RT patches
    • Docker-based microservices for modular upgrades

  • Memory Architecture

    • 16 GB DDR4 for OS and data buffering
    • 32 GB LPDDR4 for neural network inference

  • Data Buses

    • PCIe x4 connects FPGA module to SoC
    • 10 GbE switch routes sensor streams to processing cores

  • Security

    • Secure boot with RSA-2048 signed images
    • AES-256 encryption for data at rest and in transit
    • TPM module for key storage and attestation

Power management domains isolate radar, laser, and compute loads—supporting burst-mode scanning (<200 W peak) with average draw under 30 W. Battery management uses Coulomb counting and smart balancing for LiFePO₄ cells, ensuring 8+ hours of continuous operation.

Biometric Signature Database

Each pod’s on-board database supports large-scale signature management:

  • Enrollment Workflow

    • Guided UI records a 6 second sample at known GPS coordinates
    • Auto-segmentation yields ~40,000 individual beat embeddings
    • Locality-sensitive hashing compresses each to a 256-bit fingerprint

  • Indexing and Lookup

    • Primary index: hash table with 4-level bloom filters
    • Secondary index: k-d tree for embedding-space nearest neighbors

  • Matching Logic

    • Combined Hamming distance and cosine similarity scoring
    • Configurable thresholding (default: Hamming <32 bits, cosine >0.92)

  • Lifecycle Policies

    • Time-based aging demotes unused profiles after 24 h
    • Priority locking prevents removal of high-value subject signatures

Storage footprint remains under 50 MB for 1,000 active profiles, with seamless offload via encrypted USB or secure network sync.

Mechanical and Electrical Design

Ruggedization and serviceability focused design choices:

  • Structural Materials

    • Carbon-fiber composite panels over aluminum honeycomb core
    • Radar-absorbent polymer inserts for RCS reduction

  • Thermal Control

    • Heat pipes bridge hot components to ventilated chassis fins
    • Passive vents lined with hydrophobic mesh exclude water and dust

  • Vibration Isolation

    • Elastomeric mounts protect optical and radar modules from shocks up to 20 g
    • PCB mounting uses vibration-damping spacers

  • Power and I/O

    • Mil-spec circular connectors for DC in/out, Ethernet, CAN
    • USB-C port with locking latch for field data export
    • Encrypted wireless radio module supports MIMO links at 2.4/5 GHz

All assemblies undergo X-ray inspection, torque-control fastening, and conformal coating to meet MIL-STD-810G, MIL-STD-461, and IP67 sealing requirements.

Deployment and Networking

Flexible deployment options and networking stack:

  • Mount Configurations

    • Tripod adapter with leveling gimbal
    • UGV side-mount bracket with quick-release latch
    • Drone payload frame with vibration isolator

  • Mesh Networking

    • Ad-hoc routing using OLSR protocol over 5 GHz links
    • Link budgets up to 2 km per hop (20 dBi antennas)
    • End-to-end encryption with mutual authentication

  • Data Integration

    • Publish-subscribe via MQTT for C4 integration
    • REST API for track queries and signature management
    • Custom binary protocol for low-latency sensor broadcasts

Pods automatically discover peers, share track metadata, and coordinate hand-offs as targets move between fields of view—maintaining continuous tracks without central servers.

Performance Characterization

We quantify system performance against rigorous benchmarks:

  • Detection Probability vs. SNR
    ROC curves generated from 10,000 synthetic and 5,000 live test events show P_d 0.95 at SNR = 0 dB.

  • Localization Accuracy
    Cramér-Rao lower bound indicates < 0.5 m error in open terrain at 3 km range; real tests achieve RMS error of 0.8 m.

  • Identity Persistence
    Track ID-switch statistics remain under 5 % for up to 100 concurrent heartbeats in urban canyons.

  • Latency Metrics
    End-to-end processing latency averages 45 ms (±5 ms); UI update loop runs at 10 Hz with sub-20 ms rendering time.

These metrics are validated via hardware-in-the-loop setups and live field trials across desert, forest, and urban environments.

Quality Assurance and Validation

Our QA program spans virtual and physical testing:

  • Digital Twin Simulations
    High-fidelity models of sensor physics, environmental noise, and platform motion identify edge-case failures before hardware iteration.

  • Automated Component Verification
    Continuous bench testing runs 24/7 on laser line-width, radar sensitivity, and acoustic response rigs—flagging deviations beyond 3 sigma.

  • Hardware-in-Loop Regression
    Firmware updates are tested with synthetic sensor inputs to verify timing, data integrity, and algorithm stability.

  • Long-Term Soak and Stress
    Three-day climate chamber cycles (temperature, humidity, power interruptions) and MIL-STD-810G shock/vibration tests ensure reliability under mission stresses.

Defects found at any stage trigger immediate root cause analysis, corrective engineering, and regression test updates.

Path to Operational Readiness

We transition from prototypes to fielded units through iterative cycles:

  • Model-Based Calibration
    Field logs refine our simulation models, improving AI and signal-processing accuracy before the next build.

  • Cross-Domain Integration
    Synchronization with partner radar, EO/IR, and SIGINT sensors validates data fusion workflows in joint exercises.

  • Operational Workshops
    On-site sessions with end users optimize UI layouts, alert thresholds, and workflow scripts for real missions.

  • Certification Documentation
    Detailed technical manuals, training curricula, and compliance reports support rapid deployment and sustainment by partner teams.

Each iteration closes performance gaps, increases reliability, and reduces time-to-deployment.

A Prelude to the Constellation Initiative

This heartbeat tracker forms the biometric sensor layer of our upcoming Constellation Initiative—a unified platform that ingests optical, radar, chemical, RF, and biometric data into a single AI orchestrator. Constellation will provide a real-time “eye in the sky” that fuses every detection into a common operating picture, anticipates adversary intent, and guides decision-makers with actionable insights.

Message from Our Founder

At CriusCo, we don’t just reveal the invisible—we own it. Our long-range heartbeat tracker is the product of relentless, multidisciplinary R&D in physics, electronics, and AI, and there’s nothing on the planet that comes close. As we fold this powerhouse into Constellation and beyond, we dare our partners to push it into the harshest conditions—it’ll exceed expectations. Together, we’re rewriting the rulebook on life-sign sensing and showing the world how to truly perceive life itself.

Collin Baldrica
Founder & CEO, CriusCo