A.R.P. SyndicateARPSyndicate

Osprey Vision

Face & object clustering for CCTV and live drone/recon feeds.

Secure Tunnel
Encrypted Feed Ingest
RTSP + MJPEG
Any Camera or Drone Downlink
No Names
Anonymous by Default
Cloud or On-Prem
Your Choice
ILLUSTRATIVE OUTPUT — not live data
Faces — clustered, never named
cluster_000042
14 appearances · 2 cameras
cluster_000043
3 appearances · gate cam
Objects & events — detected and classified
Vehicle
silver sedan · plate logged · gate cam
Backpack
unattended 4m+ · loading dock cam
Weapon
detected · alert forwarded · rear gate
PPE compliance
hard hat missing · warehouse floor
Tailgating
2 persons, 1 badge · access cam 3
Drone sweep
214 persons counted · density: low
Runs against your own camera and drone feeds — nothing shown here is example, sample, or third-party footage. Request a live demo

Deployed on infrastructure you own or are authorized to operate — corporate sites, guarded facilities, and authorized aerial survey/recon work.

Osprey Vision connects to your RTSP or MJPEG streams — fixed cameras, NVR/DVR systems, or a drone's live recon downlink — over a secure, encrypted tunnel and clusters recurring faces and objects into opaque, internal ids. No name is ever attached unless you explicitly load your own reference watchlist for identity verification. Prefer to keep everything on your own infrastructure instead? A fully on-prem deployment is available too.
Built for lawful, authorized deployments. We do not take responsibility for any use outside the scope you are authorized to operate, and such use is entirely unintended.

What it sees

Face clusteringEvery recurring face is grouped under a stable internal id from a conservative single-face quality gate — no identity guess, just "this is the same appearance as before."
Object detection & classificationUnlike faces, objects are identified by type — vehicles, bags, PPE compliance, weapons, tailgating — since a class label ("vehicle", "hard hat missing") carries no personal identity.
Watchlist verificationOptional, separate from clustering: confirm a probe photo against a specific claimed identity using a reference set you supply — for authorized investigative correlation.

Face clustering and identity verification are deliberately separate: the general pipeline never tries to name a person, and identity verification only ever runs against a set you explicitly loaded — not the open pipeline. Object classification is a different kind of "identification" entirely — it labels a thing's type, not a person's identity.

How it works

One capture path handles both live surveillance cameras and drone downlinks, since both typically speak the same two protocols. From there, every appearance is reduced to a comparable embedding before anything gets grouped.

Secure tunnel captureA lightweight on-network agent relays your RTSP/MJPEG streams out over an encrypted tunnel — no inbound ports opened, connection always initiated from your side.
Detect & embedArcFace-family face embeddings with a quality gate on confidence and size; object detection follows the same detect-then-embed shape.
ClusterExact cosine matching against every prior appearance — not an averaged centroid — with a similarity threshold calibrated to your deployment.
CorrelateAppearance counts, first/last seen, and which camera(s) saw a cluster, so recurring presence across sites is visible without ever attaching a name.

In the on-prem deployment, all four steps run in containers on your own hardware instead — same pipeline, no tunnel involved.

Two ways to run it

Cloud (default)Your cameras and drones connect out over a secure tunnel — nothing to install beyond a lightweight agent. Detection, embedding, and clustering run in Osprey Vision's cloud; query and correlate across every site from one dashboard.
On-premFor data residency or air-gapped requirements: capture, detection, embedding, and clustering all run on your own infrastructure. Nothing leaves your network, ever.

Both run the identical pipeline — the only difference is where it executes. On-prem trades cross-site convenience for guaranteeing footage never leaves your premises.

Common questions

Does Osprey Vision identify people by name?

Not by default. The clustering engine groups appearances of the same face or object under an opaque internal id — no name is ever attached. A separate, optional watchlist verification mode can match against a reference set you supply yourself, for confirming a specific known identity — it does not run against the general clustering pipeline.

Does my camera feed leave my network?

In the cloud offering (the default), yes — your cameras' feeds are captured over an encrypted, authenticated secure tunnel and processed in Osprey Vision's cloud, which is what lets you deploy without standing up any local infrastructure and query every site from one place.

If your requirements call for footage to never leave your premises — data residency, air-gapped environments — an on-prem deployment is also available, where capture, detection, and clustering all run on your own infrastructure instead.

How does the secure tunnel work?

A lightweight agent runs on your network, connects outbound to Osprey Vision's cloud over an encrypted tunnel, and relays your configured RTSP/MJPEG camera and drone streams. No inbound ports need to be opened on your network — the connection is always initiated from your side.

What camera and feed types are supported?

Any RTSP stream or HTTP MJPEG source — the two most common transports across IP cameras, NVR/DVR systems, and drone or aerial recon downlinks. One capture path handles both.

Who is this built for?

Security operations teams and integrators deploying on cameras and airspace they own or are contractually authorized to operate — corporate sites, guarded facilities, and authorized aerial survey/recon work.

It is not a public search service, and is not intended for surveillance of individuals without a lawful basis and appropriate notice.

How accurate is the clustering?

Face clustering uses ArcFace-family embeddings with a conservative single-face quality gate on every reference sample, exact — not approximate — cosine matching, and a similarity threshold you calibrate against your own deployment's data rather than a fixed global number.

Deploy Osprey Vision

Cloud or on-prem, per-camera pricing. Tell us about your sites and camera count.

Request access Email sales