Brian Skyberg

01 Search & rescue · Computer vision

SAR Vision

AI-assisted detection that helps search-and-rescue teams review drone video for people who may need help.

PrototypeVersion v0.52
SAR Vision desktop interface showing a drone video frame of a rocky, wooded hillside with one person outlined by a detection box, a zoom preview of that person on the right, and source and detection settings on the left.

The detection interface (earlier v0.29 build) with a confirmed contact and zoom preview.

Overview

Aerial searches cover a lot of ground, and the people doing them are watching hours of video where a lost person may be only a few pixels tall. Fatigue sets in. Small details get missed.

SAR Vision is built to help with that. It runs on a field laptop, watches drone footage as it comes in, and marks anything that might be a person so the operator can take a closer look. The goal is simple: make it a little less likely that someone in the frame goes unnoticed.

General-purpose detection models are mostly trained on photos taken at ground level. From the air, people look different: foreshortened, partly hidden by trees, often lying still in clothing that blends into the terrain. SAR Vision is configured for that perspective, and for a field reality where there is often no internet connection at all.

It is a decision-support tool. The search team stays in charge of every call.

What it does today

Features

  • Real-time detection on drone video

    Analyzes live or recorded aerial footage frame by frame and highlights possible human detections as they appear.

  • Works fully offline

    All processing happens on the laptop. No cloud service, account or network connection is needed in the field.

  • Flexible video inputs

    Accepts an HDMI capture feed with automatic signal recovery, an RTMP stream, or a recorded video file. Single images and whole folders can be scanned too.

  • Tuned to surface, not filter

    Detection is configured to favor recall, so low-confidence candidates are shown for review instead of being silently discarded.

  • Confirmation and operator review

    Time-based confirmation logic separates fleeting candidates from persistent ones, with captured thumbnails and controls for the operator to verify each contact.

  • Zoom preview and run records

    A zoom panel shows each detection up close. Every run saves timestamped detection logs, thumbnails and an annotated video export for after-action review.

Real screenshots

Screenshots

Development

Current status

Prototype

Working prototype under active development. The current released baseline is v0.52; field evaluation with aircraft and operators is ongoing.

Built with

  • Python
  • PyTorch
  • Ultralytics YOLOv8
  • OpenCV

Planned next

  • Further field evaluation with real aircraft, hardware and operators
  • Specialized aerial search-and-rescue detection models, currently opt-in and under evaluation

Good to know

  • SAR Vision supports trained search personnel. It does not replace them, and every flagged contact needs human review.
  • Because it is tuned to catch as much as possible, it will flag things that turn out not to be people.
  • No detection system finds every person in every condition. Terrain, vegetation, altitude and lighting all affect results.
  • A GPU is strongly recommended. CPU-only operation works but is much slower.