Active Field Evaluation — Limited Participation

SAR Vision

Aerial Human Detection
for Search & Rescue Teams

SAR Vision processes drone video in real time and flags potential human contacts for operator review. It is designed for field conditions: no network required, no proprietary hardware, no cloud dependency. Built to support the judgment of trained SAR personnel — not to replace it.

Developed by Brian Skyberg, active SAR team member with field deployment experience.
Development Status: SAR Vision is currently undergoing structured field evaluation with select SAR teams. Access is limited while the system continues operational refinement and dataset expansion.
Real-time 1080p detection demonstrated on RTX 3060-class field laptop hardware — no proprietary equipment required
Fully offline operation — no network uplink required in the field or during active search
Recall-optimized detection — configured to surface all plausible human contacts for human review
Drone-agnostic input — HDMI capture, RTMP stream, or recorded video file from any platform
SAR_VISION ◆ INFERENCE ACTIVE 30.4 FPS
3
Contacts
RTX 3060
GPU
OFFLINE
Mode
SAR Vision in Operation
Live detection on aerial drone footage — development build
Model: Custom Fine-Tuned YOLO
Inference: Local / GPU-Accelerated
Connectivity: Offline — No Network
SAR_VISION ◆ DETECTION RECORDING
DEVELOPMENT BUILD — OPERATIONAL EVALUATION PHASE

General-Purpose Detection
Is Insufficient for
Aerial SAR Operations

Standard detection models are trained on ground-level photography from consumer and autonomous vehicle datasets. Aerial SAR presents a different visual domain — and the cost of a missed detection is not a degraded performance metric. It is a person not found.

📐

Perspective Mismatch

General models are trained on upright figures at ground level. Aerial footage shows overhead silhouettes, foreshortened limbs, and partial figures. These models were not trained to recognize subjects from above.

🌿

Terrain Camouflage

Subjects in distress are frequently stationary, wearing earth-tone clothing, and partially obscured by terrain or vegetation. Without specific aerial SAR training data, detection systems do not generalize to these conditions reliably.

📶

No Field Connectivity

Cloud-based inference depends on uplink bandwidth that does not exist across most active search areas. Any system requiring network connectivity is unsuitable for remote field deployment.

Processing Throughput

CPU-only inference cannot maintain real-time processing of 1080p drone video on standard field equipment. Processing gaps mean a subject may be in frame during an interval that was never evaluated.

⚠ The Core Asymmetry
"A false positive costs an investigation team minutes. A false negative may cost the subject their life."

Designed to prioritize recall in search-and-rescue operational contexts. This tuning strategy intentionally surfaces low-confidence detections for operator review rather than suppressing them. SAR operations invert this priority completely. Verification of a false positive is a recoverable outcome. Suppression of a valid detection is not.


SAR Vision is configured to prioritize recall over precision: all plausible contacts are surfaced for operator review, including low-confidence candidates. This is the correct engineering decision for this operational context.

Why Recall Is the
Primary Design Objective

DETECTION METRIC TRADE-OFF

Precision Secondary objective
Recall — SAR Vision (tuned) Primary objective
Recall — Base COCO (aerial) Untuned baseline
False Positive

Team investigates flagged area. No subject found. Cost: bounded time for investigation.

False Negative

Subject is in frame. System does not flag. Team continues past. Cost: potentially mission-critical.

Illustrative comparison based on internal evaluation datasets. Not independently validated.

The Model Is Calibrated
to Flag, Not Filter

Standard detection benchmarks weight precision and recall equally, or optimize toward precision because false positives carry a higher perceived cost. In SAR operations, that optimization is incorrect. SAR Vision is configured to surface all contacts the model considers plausible, including low-confidence candidates.

Confidence thresholds are configurable, but defaults are intentionally conservative. A detection at 0.45 confidence is presented to the operator for assessment rather than suppressed before review. The system aims to improve the likelihood that plausible contacts are not filtered before a human evaluates them.

This approach is reflected throughout: in training data selection, threshold configuration, and NMS parameter tuning — all oriented toward SAR context rather than general benchmark performance.

Design Principle

SAR Vision surfaces contacts. It does not adjudicate them. Every flagged detection requires human verification before any operational action. The system provides decision support; the operator retains decision authority.

What SAR Vision Does

A single-purpose detection tool: process drone video and surface human contacts to the operator in real time. Deployable on existing field hardware. No additional infrastructure required.

🎯

SAR-Specific Model

Detection model fine-tuned on annotated aerial imagery representative of search-and-rescue operational contexts — not repurposed from surveillance or autonomous vehicle training data. Trained to identify subjects from overhead perspectives in challenging terrain conditions.

🚁

Drone-Agnostic Input

Accepts HDMI capture from any drone monitor output, RTMP stream endpoints from mission planning software, or pre-recorded video files for post-flight review. No proprietary SDK required.

💻

Local GPU Processing

Runs on-device using CUDA-capable GPU hardware. A mid-range field laptop with an RTX 3060 is sufficient for real-time 1080p inference. All processing is local — no data leaves the device.

🔍

Recall-First Configuration

Confidence thresholds and post-processing parameters are configured to maximize detection coverage over precision. Low-confidence contacts are presented for operator review rather than suppressed at threshold.

📍

Operator-in-the-Loop

All flagged contacts require human confirmation. SAR Vision does not take autonomous action or determine subject status. It presents contacts for assessment; the operator decides what action, if any, to take.

🔗

Workflow Compatible

Does not replace SARTopo, CalTopo, or existing incident command tools. Adds a systematic aerial detection layer to drone operations teams are already conducting, without modifying established procedures.

System Design

Input Layer

Video Ingestion

Accepts HDMI capture (any USB capture card), RTMP endpoints from FPV or mission software, or local video files. Frame extraction is decoupled from inference to prevent throughput bottlenecks.

Pre-Processing

Frame Preparation

Frames normalized to model input dimensions. Optional adaptive contrast enhancement available for flat-lighting conditions common in overcast aerial environments.

Inference Engine

Fine-Tuned Detection Model

Custom weights derived from YOLO base architecture, fine-tuned on annotated aerial imagery representative of search-and-rescue operational contexts. Training includes overhead perspective, partial occlusion, varied terrain, and low-contrast subject presentations. GPU-accelerated via CUDA. Not standard COCO weights.

Post-Processing

Detection Output & Logging

NMS applied with SAR-tuned parameters. Contacts above configurable thresholds annotated on live feed. All events logged with timestamp and frame reference for post-mission review.

Future

SARCommand Integration

Designed as standalone module and as a detection input within the SARCommand incident management concept, currently under development.

Minimum System Requirements
GPU
REQUIREDNVIDIA CUDA-capable
RTX 3060 or equivalent
OS
Windows 10/11 · Ubuntu 20.04+
RAM
16 GB min · 32 GB recommended
Network
OPTIONALFully offline capable
Input
HDMI capture · RTMP · File
Storage
~4 GB model weights + logs
Detection Model
Architecture
YOLOv8 — custom fine-tuned
Training
Aerial SAR imagery — not base COCO
Target Class
Person — aerial/overhead perspective
Threshold
Configurable — recall-optimized default
Throughput
Real-time inference demonstrated on RTX 3060-class hardware at 1080p under field testing conditions

Designed for
Field Conditions

SAR operations don't occur in controlled environments. SAR Vision is designed to function under the constraints that actually exist during active search deployments.

📡

No Network Dependency

All inference runs locally. Model weights are bundled with the application. No license server, no cloud API, no outbound data. Operational in dead zones, canyon terrain, and remote wilderness.

🖥

Standard Hardware

Designed to run on equipment the team already carries. No specialized hardware beyond a CUDA-capable GPU. Packaging targets straightforward installation on existing field laptops.

🚁

Platform-Independent

Compatible with any drone providing a video output — DJI, Autel, Skydio, or other platforms via HDMI capture or RTMP stream. No manufacturer-specific integration required.

📁

Post-Flight Review

Recorded flight video can be processed after landing. Useful when operational conditions require full operator attention during flight, or for documentation and after-action review.

Input Source Compatibility
HDMI
USB capture card — any drone monitor output
✓ Operational
RTMP
Stream endpoint — Litchi, DJI Go, compatible software
✓ Operational
File
MP4, MOV, AVI — post-flight review
✓ Operational
Thermal
FLIR / DJI Zenmuse XT2 — RGB fusion
Planned
Multi-UAV
Parallel stream processing
Planned

SAR Vision is an additional detection layer for drone-equipped operations. It does not replace SARTopo, incident command structure, or field coordinator judgment. Teams continue operating with existing tools; SAR Vision provides systematic aerial coverage that cannot be maintained manually at scale.

Operational Testing Status

SAR Vision is in active operational evaluation. The following reflects the current testing state. No performance claims are made beyond what has been directly observed and documented.

Demonstrated to regional SAR personnel. The system was reviewed and observed by active search-and-rescue team members under structured conditions. Operational feedback from those sessions is incorporated into the development cycle.

Tested against live drone HDMI feeds. SAR Vision has been validated against real-time drone video via HDMI capture, confirming inference pipeline stability and frame throughput under field-representative conditions.

Operated under field deployment conditions. System has been run on portable field hardware, confirming offline functionality, GPU acceleration on laptop-class equipment, and detection output during active UAV flight.

Aerial SAR training data applied. Model weights reflect fine-tuning on annotated aerial imagery, demonstrating improved detection of overhead human figures compared to the untuned base model baseline.

Structured agency evaluation in planning. Formal evaluation with drone-equipped SAR units is being organized for the next development phase. Participating units will contribute structured operational feedback.

⚠   Operational Limitations

Human verification is required for all detections. No output from SAR Vision should be acted upon without evaluation by a qualified operator. The system surfaces candidates; personnel assess them.

Performance varies by terrain, lighting, and subject visibility. Detection reliability is affected by vegetation density, terrain complexity, ambient light conditions, and subject contrast against background. No system performs uniformly across all environments.

False positives are expected and by design. The recall-optimized configuration intentionally accepts a higher false positive rate. Teams should plan for flagged contacts that do not correspond to subjects on every deployment.

CUDA-capable GPU is a deployment prerequisite. CPU-only hardware is not recommended for real-time deployment. This requirement must be confirmed before evaluation planning begins.

Active development build. SAR Vision is under continuous development. Evaluation units should expect iterative updates and are expected to provide structured operational feedback as part of participation.

Annotated Detection Scenarios

Representative detection scenarios illustrating the types of contacts SAR Vision surfaces, including the reasoning behind each flag. Annotated screenshots from active field evaluation will replace these placeholders as testing progresses.

SCENARIO 01
[ Awaiting field capture ]
Subject at Brushline
0.87 CONF

Overhead pass at approximately 80m AGL. Subject partially obscured by light vegetation at terrain edge. Untuned baseline model did not flag this frame under default confidence thresholds.

Flagged: Upper-body silhouette visible through partial canopy — consistent with SAR training examples for prone or seated subject positions at terrain margins
SCENARIO 02
[ Awaiting field capture ]
Low-Contrast Rocky Terrain
0.61 CONF

Subject in dark clothing against dry granite and scrub. Overcast conditions, flat ambient light. Low-confidence contact surfaced under recall-first configuration rather than suppressed.

Flagged: Edge profile and limb geometry consistent with human figure — presented for verification under recall-first threshold; below standard precision-optimized cutoff
SCENARIO 03
[ Awaiting field capture ]
Prone Subject, Open Meadow
0.93 CONF

Subject in prone position in open terrain. High overhead angle produced atypical aspect ratio. Untuned baseline model did not flag this frame under default confidence thresholds; SAR-specific model flagged correctly.

Flagged: Horizontal figure profile and color contrast against grass — pattern consistent with prone human subjects as represented in SAR aerial training data
⚠ IMPLEMENTATION NOTE — Scenario images above are representative placeholders. To add real captures: replace each .det-placeholder div with a standard <img> tag. Caption, confidence badge, and flag-reason structure are in place and require no additional HTML changes.

Example frames shown are representative field test imagery. Performance varies based on terrain, lighting, altitude, and subject visibility.

Who This Tool Is For

SAR Vision was built for a specific operational context. Understanding where it fits — and where it does not — is part of evaluating whether it is appropriate for your unit.

✓   Appropriate For
Active SAR teams with drone programs

Units conducting UAS-assisted searches who need systematic coverage of aerial video beyond what manual operator review can sustain.

Units in low-connectivity environments

Teams whose search areas lack reliable cellular or satellite uplink — remote wilderness, canyon terrain, or backcountry without communications infrastructure.

Teams prioritizing recall over automation

Personnel who want to increase detection coverage without delegating subject identification to the system — operators remain in the loop on every flagged contact.

County SAR, sheriff aerial units, Civil Air Patrol

Government and accredited volunteer units with structured operational procedures who can integrate detection assistance into existing field workflows.

✕   Not Designed For
Fully autonomous detection systems

SAR Vision requires active operator oversight. It is not architected for unattended or autonomous drone patrol without human monitoring.

Consumer or hobby drone users

Designed for organized SAR teams with formal operational structures and trained personnel. Not intended for personal or recreational aerial use.

Cloud-dependent deployment models

SAR Vision runs locally and does not transmit data externally. Teams requiring centralized remote processing infrastructure should evaluate other solutions.

Autonomous subject location reporting

The system does not provide autonomous GPS coordinates, subject tagging, or automatic dispatch triggers. All detections require human review and interpretation before any action.

SAR Vision Is Decision-Support Software

All detections require human verification. No detection output from SAR Vision should be acted upon without evaluation by a qualified operator. The system surfaces candidates; trained personnel assess them.

SAR Vision does not replace operator judgment. Aerial observation, search pattern planning, and subject determination remain under the authority of qualified SAR personnel. The system is an analytical aid, not a decision authority.

It is not an autonomous search system. SAR Vision does not direct aircraft, prioritize search sectors, or make resource allocation decisions. These functions remain entirely with incident command and field personnel.

Detection rates are environment-dependent. No threshold guarantees that all subjects present in video will be flagged. System performance should be understood as probabilistic assistance, not exhaustive coverage.

Planned Development

SAR Vision is under active development. The current build addresses human detection from RGB aerial video. Subsequent phases are planned based on operational priority and feedback from evaluation units.

Active — Current Build

Phase 1
Human Detection

  • Aerial person detection — fine-tuned model
  • HDMI, RTMP, and file input sources
  • Local GPU-accelerated inference
  • Configurable confidence thresholds
  • Detection logging and timestamping
  • Field laptop deployment packaging
  • Recall-optimized configuration
Phase 2 — In Planning

Phase 2
Thermal & Animal Detection

  • Thermal input — FLIR, DJI Zenmuse XT2
  • RGB + thermal multi-modal detection
  • Animal detection class for wildlife SAR
  • Heat signature flagging
  • Night operational capability via thermal
  • Expanded training data — thermal domain
Phase 3 — Concept

Phase 3
Multi-Drone & SARCommand

  • Parallel stream processing — multi-UAV
  • Shared detection view across operator stations
  • SARCommand integration — detections to sectors
  • KML / GeoJSON export for SARTopo
  • Mission replay and detection audit trail
  • Team-facing detection notification interface

Roadmap priorities are shaped directly by feedback from evaluation units. If your operational context involves specific terrain types, detection challenges, or equipment constraints not addressed here, that input is sought and directly informs development sequencing.

Submit Operational Inquiry

Structured Field Evaluation

SAR Vision is currently undergoing structured field evaluation with select SAR teams. Participation is limited while the system continues refinement, operational validation, and training dataset expansion.

This is a collaborative development phase. Participating units are expected to contribute observations on detection performance, false positive rates, field usability, and integration with existing procedures. That feedback directly determines development priorities.

Participation is coordinated directly with evaluation units, including structured check-ins, deployment guidance, and feedback review after operational use.

Evaluation Criteria
  • Active SAR team with an established UAS program and regular drone deployment on operations
  • Operations in environments with limited or no cellular or satellite uplink
  • Access to a CUDA-capable GPU laptop or workstation for evaluation
  • Willingness to provide structured feedback following evaluation flights
  • Unit operates under formal SAR jurisdiction or accreditation (county, state, federal, or recognized volunteer)

Operational Inquiry

Field evaluation participation — active SAR units only

Submissions are routed through your default email client for direct evaluation coordination follow-up.

Operational Boundaries