Stefan
SPECTRA analyses a controlled night scene with a person and a van on commercial premises.
SPECTRAEvidence vision

Multispectral security / evidence AI

SPECTRA

A working browser proof that translates camera image into demonstrable detection, measurable quality limits and human verification.

Working evidence-vision proof

The case in 30 seconds

SPECTRA

Problem
A working browser proof that translates camera image into demonstrable detection, measurable quality limits and human verification.
Solution
From your own camera image to local detection, measurable image quality, level of evidence and exportable audit record.
Role and focus
Night vision, evidence AI
Status
Working evidence-vision proof
Result
Case, product site, live lab

Live lab / try it yourself

Test what camera image really proves

Use your own or sample image, compare detection with image quality and see when the system shows evidence or deliberately does not draw a conclusion.

Open the vision lab View the story
01 / InputCamera image and quality limitYou can change or choose this yourself
02 / SystemLocal vision AI and verificationComputer vision · ONNX inference · Evidence scoring
03 / EvidenceA measurable level of evidenceThis makes the operation visible
InputOwn camera image
AnalysisLocal object detection
EvidenceQuality + source framework
Safety boundaryDon't make up an identity

The design question

When does image become evidence?

A sharp live feed is not automatically useful for a safety decision. The task determines what counts: detecting someone requires a different image than recognizing clothing or confirming an identity.

01

Detect

Is there demonstrably a person or vehicle present?

lowest image demand
02

Observe

Are direction, action and context visible?

more pixels + time
03

Recognize

Are distinguishing features reliably observable?

detail + good lighting
04

Identify

Can an identity be confirmed with an appropriate method and basis?

separate validation

Specialist layer · real model run

Ten difficult goals

Grounding DINO Tiny only received text prompts per scene. Ten predetermined target questions were located in this manually controlled lab set. The specialist route on these images proves that, not an overall accuracy score.

Grounding DINO locates a fat bike in a pedestrian zone
Open vocabularyFatbike · 75.7%
Grounding DINO locates a mobility scooter on a square
Specialist classMobility scooter · 77.5%
Grounding DINO distinguishes a motorcycle and a van in one scene
Two vehicle typesMotorbike 93.8% · van 86.6%
Grounding DINO distinguishes between a stroller and a shopping cart in a supermarket
Context objectsStroller 75.1% · shopping cart 51.0%

Grounding DINO Tiny · box threshold 0.24 · text threshold 0.20 · executed locally · images remain internal. The outcome has been manually checked and remains explicitly lab-proof.

Technical status · fairly defined

From live model to production path

Not every part is in the same phase.SPECTRAexplicitly shows what currently works in the browser, what has been proven on a fixed lab set and which technology only forms the next production step.

Live nowYOLO26s · ONNX Runtime Web

Basic model in the browser

Local object detection at 960 pixels with WebGPU and WASM fallback, plus quality measurement and exportable audit record.

Proven on fixed lab setGrounding DINO Tiny

Open vocabulary as a specialist

Text-driven localization of more difficult mobility and context objects with explicit thresholding and human control.

Production pathRF-DETR · ByteTrack · BoT-SORT

Add your own classes and time

First fine-tune your own labeled vehicle classes; then follow route and duration of stay. Distance and speed require separate camera calibration.

Nighttime sample image for theSPECTRAbrowser analysis
person91% reviewed

Live in the browser

Test your own frame

Choose a JPG, PNG or WebP. The model runs locally in the browser, drawing real detections and simultaneously assessing light, contrast, sharpness and target size.

  • No upload to a server
  • Original image remains visible
  • Detection is not identity
  • Insufficient evidence remains a valid outcome
Open the live lab →

Design principles

Evidence before effect

SPECTRAdoes not make uncertainty more beautiful, but useful. Every conclusion remains linked to source pixels, measurement values ​​and a human follow-up step.

Bronvast

No generated face

Clarification may help you look. Generated details or hallucinations never become forensic evidence.

Measurable

Task-oriented quality gate

The system reports why a frame is suitable for detection but not for recognition.

Human

Verification remains visible

An operator receives the original image, marker, score, boundary and recommended next step.

Complete product experience

From frame to decision

First see why the claim persists or immediately run your own camera image through the local analysis.

OpenSPECTRAStart live analysis

More intelligent cases

Other systems with demonstrable value

From community and gameplay to brand and hospitality systems.

AELINanalyzes a gentle morning walk with the real 33-point body pose and risk framework. Care vision / human verification AELIN From fixed space camera to explainable pose, risk zone and a quiet human verification route. MISEforecast with expected covers, turnover band, capacity development and team advice for a catering day. Hospitality intelligence/operation MISE A reproducible hospitality-proof with 546 synthetic days, 65 unseen test days, table model, staff scheduling and fail-closed data protection. AURELIA 67anchored in clear tropical water. Digital twin / luxury experience AURELIA 67 A private digital twin that connects form, materials and ownership experience. Stradviso Decision Studioon tablet, desktop and phone. AI product/decision support Stradviso Decision Studio Controlled AI decision system that converts customer questions into traceable options, human approval and audit trail.