SkyWatch — Aerial Search & Rescue Vision Platform

When someone goes missing in 10,000 acres of forest, every minute costs survivability. SkyWatch turns a search drone into an autonomous spotter — scanning terrain with fused thermal and visual AI and surfacing human-shaped heat signatures to incident command the instant they appear.

Computer VisionPyTorchYOLOThermal/RGB FusionEdge AI (Jetson)ROSWebRTCMapboxNode.js

Industry

Public Safety / SAR

Scale

Large · Flagship system

Status

Field-deployed

// Problem

The challenge

Ground search teams cover wilderness slowly and exhaustively, and conventional drone footage still depends on a human operator squinting at a tiny screen for hours. Fatigue causes missed detections; thermal cameras alone throw false positives on sun-warmed rock and wildlife. The county SAR unit needed to search far more ground per flight hour without adding analysts they didn't have.

// Solution

What we built

An end-to-end aerial vision platform that runs detection on the aircraft, so findings reach searchers in real time over degraded rural links.

Dual-sensor capture (radiometric thermal + 4K RGB) with frame-level geo-registration to ground coordinates
On-board inference on an NVIDIA Jetson — a YOLO-based detector fine-tuned on aerial human-form and heat-signature data, plus a fusion stage that cross-confirms thermal hits against visual cues to kill false positives
Autonomous "lawnmower" survey routing over a drawn search polygon, with terrain-aware altitude hold
A live incident-command web dashboard: detections drop onto a map as pins with a confidence score, snapshot, and one-tap "vector a team here"
Every flight archived for review, retraining, and after-action reporting

// Architecture

How it works

The dricraft streams compressed detections (not raw video) over WebRTC + a store-and-forward buffer, so a finding survives a dropped link and re-sends when signal returns. A Node.js command server fans detections to every connected device, while full-resolution imagery syncs to S3-compatible storage when the aircraft lands. The whole model pipeline is containerized so a new detector version can be pushed to the fleet without re-flashing hardware.

// Outcome

Results

~6× more ground covered per flight hour than manual spotting, with detections triaged for a human in the loop rather than replacing them
False-positive rate cut dramatically by requiring thermal+visual agreement before a pin is raised
Detections reach the incident commander in seconds, geo-located to within a few meters
A reusable airborne-vision foundation now extended to wildfire hot-spot mapping and flood-damage assessment

// Have a similar problem?

Let's talk about what you need built.

Custom-engineered solutions — no generic platforms, no compromises.

Start a Project →