Nexma

Object Detection

Find & classify objects at scale

Object detection is Nexma's perception capability — finding and classifying objects in imagery and video at scale, across land, sea, and air. Vehicles, vessels, aircraft, infrastructure, terrain: detected, classified, tracked across frames, and written straight into the world model as typed entities. Imagery and video arrive faster than any team can watch; the InsightEngine turns that flood into something you can query and act on.

It is the same operating system — DataStore, Ontology, Jax, and the engines — with the perception layer doing the watching.

What you can do

  • Detect and classify at scale. Production-grade computer vision finds objects across overhead imagery and full-motion video, with a confidence score on every hit.
  • Track over time. Object identity is held across frames and passes, so a vessel is followed through time rather than re-detected as a stranger.
  • Fuse many sensors. Satellite, drone video, and ground sensors merge into one deduplicated, georeferenced picture — no single-source blind spots.
  • Detect change. Automated before-and-after analysis of any area flags new construction, movement, and encroachment without frame-by-frame review.
  • Index and query. Every detection is geolocated and indexed in the DataStore, queryable by area, time, class, or confidence, with a full audit trail.

Core concepts

A bounding box in a viewer is not actionable. The point of detection in Nexma is to make perception part of the world model — so a detection is a typed entity that can trigger a workflow, not a pixel rectangle that sits in a tab.

  • Detections become entities. Results land in the DataStore typed by the active Ontology, carrying model, confidence, and source. A detection can launch an alert, a task, or a dispatch.
  • Tracking gives the picture memory. Holding identity across frames means the operational picture knows what has appeared, moved, or changed since the last pass — not just what is in the current frame.
  • Fusion gives one truth. Hits from every sensor land on the same map, deduplicated and georeferenced, so teams reason from one shared picture instead of a stack of disconnected viewers.
A detection that lives in the world model can trigger an agentic workflow automatically. A detection in a standalone viewer cannot. That difference is the whole capability.

How it works

Point the perception layer at any overhead or motion feed. It detects, tracks, fuses, and grounds — turning raw pixels into entities the rest of the platform understands.

1Feed: full-motion video over an area of interest 2 1. Detect classify every object in frame, with confidence 3 2. Track lock a track onto each one; hold identity across frames 4 3. Fuse merge with EO, SAR, and ground-sensor hits → one picture 5 4. Ground write typed entities to the DataStore (model + confidence + source) 6 5. Act a detection can trigger an AgentEngine workflow

Models watch continuously, so events surface automatically instead of waiting in an analyst's backlog.

Perception across modalities

StageWhat it covers
PerceiveDetection across EO, SAR, and full-motion video; change detection; multi-modal input
Track and fuseTemporal tracking across frames and passes; deduplicated, georeferenced fusion; confidence and provenance per hit
Ground and actTyped entities in the DataStore; workflow triggers; full history retained for replay

Example

A port authority needs continuous awareness of vessel activity in a restricted anchorage.

  1. Point the perception layer at the live satellite and coastal-camera feeds over the anchorage.
  2. Detection classifies every vessel in frame and locks a track onto each; AIS hits are fused so dark vessels stand out from broadcasting ones.
  3. Each detection is written to the DataStore as a typed entity with model, confidence, and source.
  4. A vessel entering the restricted zone without an AIS broadcast crosses a threshold, and the detection triggers an AgentEngine workflow that alerts the watch.
  5. Later, the team queries every detection in the zone over the past week by class and confidence to build a pattern-of-life view.
Swap the Ontology and the same detect-track-fuse-ground loop serves construction-progress monitoring, wildlife survey, or asset inventory. The perception is generic; the meaning comes from the world model.

Where to go next