Ring Compass: Touchless Spatial Safety & Biometrics

Transforms Ring cameras into a touchless biometric radar for seniors: detects wandering and falls, extracts vitals through the air, and triggers Alexa ambient pathway lighting with zero wearables.

Comment

Inspiration

Across the globe, over 55 million people live with Alzheimer’s disease or other forms of dementia—a figure projected to skyrocket to 139 million by 2050. Among the most harrowing and dangerous symptoms of cognitive decline is nocturnal disorientation and sleep-wandering.

When an elderly individual wakes in the dead of night, disorientation combines with sleep inertia. In their confusion, they step out of bed into darkened hallways, lose their spatial bearings, and begin pacing erratically. Within minutes, acute confusion frequently culminates in a catastrophic fall. Unintentional falls are the leading cause of fatal and non-fatal injuries among older adults; for a senior with osteoporosis or cognitive impairment, a single midnight fall often marks the permanent loss of independent living.

The Breakdown of Existing Solutions Families and professional caregivers are desperate for reliable monitoring, yet existing market offerings fail at the fundamental point of care:

The Fallacy of Wearables: Wearables (smartwatches, wristbands, emergency pendant lanyards) assume voluntary human cooperation. A disoriented senior suffering from dementia routinely removes "foreign" objects from their wrists, forgets to put on a pendant before sleeping, or fails to recharge device batteries. When a crisis occurs at 2:30 AM, the pendant is invariably sitting on a nightstand. Pressure Mats and Floor Sensors: Bedside pressure mats suffer from catastrophic false-alarm rates. Simply sitting up to reach for a glass of water trips the buzzer, jarring the senior into panic and causing caregivers to suffer severe alarm fatigue—leading them to eventually unplug the system entirely. Passive Security Cameras: Traditional smart home cameras are purely passive observation tools. They stream unanalyzed video to cloud servers, raising massive privacy and dignity concerns in private bedrooms and hallways. Worse, their alerts consist of blunt, contextual-void push notifications ("Motion detected at 2:14 AM"). When a caregiver is awakened by a ping in the middle of the night, their bleary-eyed brain desperately needs answers to The Four Critical Questions in Order:

01: Is it real? (Filtering out a roaming house pet, shadows from car headlights, or an oscillating fan). 02: What changed? (Distinguishing whether the senior is walking purposefully to the bathroom or pacing in an erratic, disoriented loop). 03: What do I do? (Intervening immediately before the senior reaches an exterior exit door or collapses). 04: Can I go back to sleep? (Verifying that the senior has been guided safely back to bed and their physiological vitals are homeostatic). Current commercial tools answer only the first question, and even then, poorly.

We conceived Ring Compass to completely redefine senior safety. Our core thesis was simple yet radical: transform commodity optical camera lenses into an intelligent, touchless biometric radar and spatial safety system that operates entirely on local edge silicon with zero wearable hardware, zero cloud video leakage, and autonomous physical de-escalation.

What it does

Ring Compass is an end-to-end ambient spatial computing platform. By pairing local computer vision with a deterministic safety engine and edge IoT hardware, it continuously protects vulnerable seniors without requiring them to wear, touch, or configure a single physical sensor.

  1. Touchless Through-the-Lens Biometric Telemetry (EVM) Using Eulerian Video Magnification principles, Ring Compass isolates the micro-vascular capillary bed of the senior's forehead. As the heart pumps blood, minute, sub-visual fluctuations in arterial skin color occur. The system isolates these micro-luminosity signals across temporal frames, measuring:

Heart Rate (BPM): Touchless optical pulse rate extraction between 45 and 180 BPM. Respiration Rate (RPM): Thoracic cavity excursion tracking between 10 and 32 breaths per minute. Forehead Optical Coherence: Signal-to-noise quality metric ensuring valid human contact before vitals are certified.

  1. 2D Spatial Centroid Tracking & Kinematic Dispersion Ring Compass projects the senior's physical floor coordinates onto a 2D spatial coordinate plane in real time:

Kinematic Trajectory Vectors: Generates a real-time trail of the subject’s center of mass $(x_t, y_t)$ across calibrated home zones (Interior Corridor, Entrance Threshold, Secured Exterior Doorway, Porch Landing). Disorientation Cadence Index (DCI): A proprietary mathematical ratio comparing total path distance traversed against net spatial displacement. Purposeful walking to the kitchen produces a low DCI ($\approx 1.0\text{x}$), whereas agitated wandering or pacing loops generate elevated DCI spikes ($> 3.0\text{x}$).

  1. The 6-Stage Deterministic Safety Gate The platform runs a deterministic decision gate every 16 milliseconds to classify household events:

Nominal Transit: Directed movement with healthy biometrics; logged silently without disturbing the household. Profile A (Agitation / Wandering): Triggered when the senior enters the egress threshold with $\text{DCI} > 3.0\text{x}$ and elevated vitals ($\text{HR} > 95\text{ BPM}$). Profile B (Vertical Fall / Rapid Collapse): Triggered when the senior's vertical centroid drop velocity exceeds $42\text{ px/s}$ and the bounding box aspect ratio abruptly collapses from vertical ($> 1.8$) to horizontal ($< 0.8$).

  1. Autonomous Alexa Connected Physical De-escalation Instead of triggering shrill, panic-inducing sirens, Ring Compass initiates gentle, therapeutic de-escalation:

Simulated Amber Pathway Lighting: The local system illuminates amber edge lighting along floor baseboards. Amber light ($590\text{ nm}$) stimulates the retino-hypothalamic tract without inhibiting melatonin production, visually anchoring the senior and showing them the path back to bed. 432 Hz Resonant Acoustic Grounding: An edge audio transducer broadcasts a warm 432 Hz grounding alert chime, followed by synthetic voice guidance ("Everything is safe. Follow the amber pathway back to bed").

  1. Verified Caregiver Emergency Dispatch If a Profile B fall is triaged, Ring Compass bypasses app-based push notifications (which are frequently silenced at night) and immediately dispatches an automated Caregiver Emergency Alert Email with:

High-resolution JPEG snapshot of the fall event with tracked centroid reticle. Pre-fall and post-fall vital signs (BPM, RPM). Exact timestamp and zone of impact. Direct action links to verify household safety.

How we built it

Ring Compass was engineered from the ground up as a high-performance local edge pipeline:

  1. Eulerian Video Magnification (EVM) Mathematical Core To extract blood pulse volume changes from standard RGB optical frames without touching the skin, we implemented a temporal Eulerian magnification pipeline:

Forehead ROI Isolation: Given input frame $F_t$, face detection extracts the upper-facial bounding box containing the forehead capillary network: $$\text{ROI}t = F_t[y{\text{forehead}} : y_{\text{forehead}} + h, ; x_{\text{forehead}} : x_{\text{forehead}} + w]$$

Spatial Decomposition: To suppress high-frequency spatial camera noise while retaining diffuse color shifts, spatial smoothing averages the green channel (where hemoglobin oxygen absorption is maximal): $$s(t) = \frac{1}{|\text{ROI}t|} \sum{(x,y) \in \text{ROI}_t} G_t(x, y)$$

Temporal Bandpass Filtering: The raw spatial mean signal $s(t)$ is passed through a zero-phase 4th-order Butterworth bandpass filter centered on physiological cardiac pulse frequencies: $$[f_{\text{low}}, f_{\text{high}}] = [0.75\text{ Hz}, 3.0\text{ Hz}] \quad \iff \quad [45\text{ BPM}, 180\text{ BPM}]$$

The filtered cardiac waveform $y_{\text{cardiac}}(t)$ is obtained via the transfer function: $$H(z) = \frac{\sum_{k=0}^{N} b_k z^{-k}}{1 + \sum_{k=1}^{N} a_k z^{-k}}$$

Spectral Estimation via Fast Fourier Transform: The dominant frequency $f_{\text{peak}}$ is isolated by computing the discrete Fourier transform over a moving window of $N = 128$ frames at sample rate $f_s = 30\text{ FPS}$: $$S(f) = \sum_{n=0}^{N-1} y_{\text{cardiac}}[n] e^{-j 2\pi f n / f_s}$$ $$\text{BPM} = 60 \times \arg\max_{f \in [0.75, 3.0]} |S(f)|$$

A parallel filter tuned to $[0.16\text{ Hz}, 0.50\text{ Hz}]$ ($10\text{ to }30\text{ RPM}$) monitors thoracic expansion and contraction to extract the Respiration Rate (RPM).

  1. Spatial Kinematics & Disorientation Cadence Index (DCI) To track spatial disorientation, the system extracts the subject's center-of-mass centroid $C_t = (x_t, y_t)$ across time. Over a continuous temporal window of $W = 30$ frames (1.0 second of motion), we calculate:

Cumulative Path Traversed ($L$): $$L = \sum_{t=1}^{W} \sqrt{(x_t - x_{t-1})^2 + (y_t - y_{t-1})^2}$$

Net Euclidean Displacement ($\Delta D$): $$\Delta D = \sqrt{(x_W - x_0)^2 + (y_W - y_0)^2}$$

Disorientation Cadence Index ($\text{DCI}$): $$\text{DCI} = \frac{L}{\Delta D + \epsilon}$$ (where $\epsilon = 0.001$ prevents division by zero).

Physical Interpretation: Directed Transit ($\text{DCI} \in [1.0, 1.8]$): The individual walks in a straight line from their bedroom to the bathroom ($L \approx \Delta D$). Verdict: Nominal Transit. Agitated Wandering ($\text{DCI} > 3.0$): The individual loops back and forth in front of an exterior doorway or pacing in circles ($L \gg \Delta D$). Verdict: Profile A (Wander Crisis).

  1. Fall & Collapse Velocity Vector Mechanics A fall is fundamentally an instantaneous kinetic state transition. We compute the vertical drop velocity $v_y(t)$ and bounding box aspect ratio $\text{AR}(t)$:

$$v_y(t) = \frac{y_t - y_{t-k}}{k \cdot \Delta t}$$ $$\text{AR}(t) = \frac{\text{Height}_t}{\text{Width}_t}$$

When an individual collapses:

$v_y(t) \ge 42.0\text{ pixels/sec}$ (rapid downward acceleration). $\frac{d}{dt}\text{AR}(t) < -1.2\text{ s}^{-1}$ (instantaneous transition from an upright vertical posture to a horizontal floor posture). When both criteria hold concurrently for 3 consecutive frames, Profile B triggers deterministically in $<42\text{ milliseconds}$.

  1. The Local MCP Spatial Server & Alexa Hardware Architecture To completely eliminate external cloud dependencies and AWS hosting costs during evaluation, we designed our architecture around the Model Context Protocol (MCP):

FastAPI Backend (Python 3.14): Serves high-speed JSON-RPC tool endpoints over port 8000. In-Memory Biometric Telemetry Bus: Computes and caches vital signs, centroid trails, and alert states with sub-millisecond dispatch times. ESP32 Edge Microcontroller Node: Connects over local WiFi to the MCP Server. GPIO 4 (Simulated Smart Bulb): Drives a high-output amber LED strip providing $590\text{ nm}$ ambient pathway illumination. GPIO 15 (Simulated Smart Speaker): Drives a 432 Hz acoustic piezo transducer generating harmonic alert tones.

  1. Frontend Telemetry Dashboard Built using React 19, TypeScript, and Vite, styled with Vanilla CSS & TailwindCSS, the web console operates as a mission-control telemetry station:

5 Core Navigation Tabs: Overview, Motion Analysis, Home & Voice, Safety Engine, and System & Events. Dual Live Streams: 60 FPS feeds with real-time target reticles and coordinate overlays. Interactive 2D Spatial Centroid Plot: Dual phase-plane plots illustrating pacing loops against room zone thresholds. Dynamic Recharts AreaChart: Live Disorientation Cadence Index stream charting real-time spikes against the $3.0\text{x}$ warning threshold. Push-to-Talk & Intercom Guidance: Browser-based audio controls simulating Alexa companion speech broadcasts.

Challenges we ran into

  1. Sub-Visual Luminance Shifts vs. Ambient Lighting Noise The Challenge: Eulerian pulse amplification amplifies everything in the frequency band—including flickering AC light fixtures, shadows cast by outdoor branches, and subtle head nodding. In early prototypes, a senior turning their head caused the estimated heart rate to artificially spike to 170 BPM.

The Solution: We implemented an adaptive region-of-interest stabilization algorithm. Using 2D affine transformations, the forehead patch is dynamically locked relative to facial landmark contours before green-channel extraction. Furthermore, we implemented a temporal optical coherence threshold: if the spatial variance across the forehead patch exceeds $15%$, the frame is discarded as motion artifact rather than pulse signal.

  1. Eliminating ML Hallucinations in Life-Safety Triage The Challenge: Modern multimodal LLMs (e.g., GPT-4V or Gemini Vision) can reason about complex video scenes, but their inference latency is measured in seconds ($1.5\text{s} - 4.0\text{s}$), and they suffer from non-deterministic hallucinations. In an emergency where an 82-year-old collapses, a 4-second delay or a hallucinated false-negative is catastrophic.

The Solution: We deliberately decoupled scene description from safety triage. While generative models are excellent for companion speech, the safety core of Ring Compass is a strictly deterministic, 6-stage mathematical state machine. By relying on geometric and kinematic equations ($v_y \ge 42\text{ px/s}$, $\text{DCI} \ge 3.0$), the triage verdict executes deterministically in under 42 milliseconds on local CPU cycles alone.

  1. Cloud Developer Sandboxing & Zero-Cost Architecture (The Friction Log) The Challenge: The commercial Amazon Alexa Plus developer console restricts beta developer SDK access for external endpoint binding and requires costly AWS cloud infrastructure (Lambda, DynamoDB, API Gateway, S3) that can quickly generate unexpected developer bills during prototyping.

The Solution: We authored a comprehensive Friction Log documenting this barrier and architected a completely sandboxed local Model Context Protocol (MCP) gateway. By wrapping our OpenCV video engine in JSON-RPC schemas matching Alexa Smart Home Skill conventions, we built an end-to-end working system that costs $0.00 in cloud bills while remaining 100% compliant and ready for production Alexa Plus deployment.

  1. Frame-Accurate Hardware/Video Audio Synchronization The Challenge: When recording physical hardware validation clips (demonstrating the ESP32 amber LED and acoustic buzzer) alongside synthetic voiceovers, stream-level video concatenation created a subtle $+0.76\text{s}$ audio drift. This caused the hardware buzzer sound to bleed into subsequent narration segments.

The Solution: We re-engineered our production pipeline using FFmpeg’s sample-accurate filter_complex concat ([0:v][0:a]...concat=n=13:v=1:a=1), enforcing precise microsecond boundaries on both video frames and audio samples. We also programmed a clean $1.2\text{s}$ silence tail on the hardware clip, guaranteeing that the narrator speaks over crystal-clear silence.

Accomplishments that we're proud of

  1. 100% Touchless Biometrics Through Ordinary Camera Lenses: Proving that heart rate and respiration can be reliably estimated across optical room cameras without attaching a single electrode, watch, or strap to a senior.

  2. True On-Device Privacy Architecture: Solving the senior care privacy dilemma. Raw video streams are processed volatilely in edge RAM; no visual footage is ever transmitted to, stored in, or monetized by third-party cloud servers.

  3. Sub-42 Millisecond Triage Latency: Achieving deterministic, life-saving triage verdicts between wandering and fall collapse in less than half the time of a human blink.

  4. End-to-End Cyber-Physical Loop: Moving beyond software mockups to build a fully functional physical hardware bridge with an ESP32 node, demonstrating real amber pathway lighting illumination and acoustic grounding tones.

  5. Verified Autonomous Emergency Dispatch: Demonstrating automated generation and delivery of high-priority emergency emails with forensic photographic evidence and biometric readouts to designated caregiver inboxes.

  6. Strict Production Discipline: Delivering a master video under 3.25 minutes ($187.06\text{s}$), accompanied by frame-accurate subtitles and a clean, open-source codebase purged of all media bloat.

What we learned

  1. Kinematics Beat Brute-Force Neural Networks: Heavyweight 3D pose-estimation models consume immense GPU resources and struggle with occlusion. In contrast, tracking the 2D centroid path integral ($L$) and displacement ($\Delta D$) provides a robust, lightweight indicator of cognitive distress that runs effortlessly on low-power edge processors.

  2. The Clinical Psychology of Senior De-escalation: Blaring sirens and harsh white spotlights trigger the sympathetic nervous system, inducing panic and increasing fall probability. Ambient amber lighting ($590\text{ nm}$) and low-frequency harmonic chimes ($432\text{ Hz}$) engage the parasympathetic nervous system, gently guiding confused seniors toward safety.

  3. The Power of the Model Context Protocol (MCP) for Edge AI: MCP is traditionally viewed as a protocol for cloud-hosted AI agents. We demonstrated that MCP can be deployed as an edge abstraction layer, allowing local AI controllers to orchestrate cameras, biometrics, and physical IoT devices with zero cloud lock-in.

What's next for Untitled

  1. Official Alexa Plus Smart Home Skill Certification: Transitioning our sandboxed local MCP server into an official production Alexa Plus Skill, enabling natural language inquiries through Echo Show devices ("Alexa, what was Mom's spatial trajectory overnight?").

  2. Whole-Home Multi-Camera Spatial Handoff Mesh: Expanding our 2D centroid tracker into a distributed spatial mesh that seamlessly tracks senior transitions between multiple Ring Indoor Cams and Doorbell nodes as they move across rooms.

  3. Clinical Memory-Care Pilot Trials: Partnering with assisted-living memory-care facilities to conduct clinical validation trials, comparing Ring Compass's fall-prevention efficacy and sleep disturbance rates against traditional pressure mats.

  4. Edge TPU & NPU Hardware Acceleration: Porting Eulerian video magnification filters and OpenCV pipelines to low-cost neural accelerators (such as the Google Coral TPU or Hailo-8), enabling Ring Compass to run directly inside future Ring smart cameras as native on-chip firmware.

Built With

Share this project:

Updates

Submission history