DEFENSAR Technology

We analyse video where it is captured.

We develop the hardware and software for edge video analytics. Detection, tracking and rules run next to the cameras, and only events that need a response travel any further.

Get in touch See how it works

3 m 25° 2.7 m 86 m Two systems on one site. The perimeter camera (4 Mpx, 4 mm lens, 3 m up, tilted 25°) does not see a pie slice: its footprint starts 2.7 m from the mast, leaving a blind spot at its foot, and confidence falls with distance. A person who crosses the fence raises an alarm. At the passive level crossing, a solar-powered mast carries a camera and a radar. Every car must stop at the line; one that does not is recorded. When the radar picks up a train 300 m out, the STOP signs light and cars wait. A trackside indicator tells the driver whether the crossing is clear: if a vehicle or a person stands on it, the train brakes and moves on only once the crossing is empty. Seagulls, animals and passing traffic are detected and ignored. Move your cursor into the frame and the device will start tracking you. Drag a finger across the frame and the device will start tracking you.

From stream to alarm in a single device

Every camera goes through the same pipeline. Everything except the first and the last step happens locally, on the device next to the recorder, so no continuous stream leaves the site.

  1. Camera stream

    RTSP from existing cameras or the recorder, H.264 and H.265. Cameras are discovered over ONVIF, whatever the brand.

  2. Hardware decoding

    A dedicated video unit decodes and scales the image. The CPU stays free, and one device handles many cameras at once.

  3. Detection on the NPU

    Our own models find people, vehicles and animals, including at night under infrared. The far part of the frame is analysed at full resolution.

  4. Object tracking

    Every object gets its own track, so rules judge behaviour over time rather than a single frame.

  5. Rules and arming

    Zones, directional lines and loitering time. The device arms together with the alarm panel’s partitions or on a schedule.

  6. Alarm to the station

    The event and a video clip travel over an authenticated two-way channel to the monitoring station or VMS. The device reports in every minute and flags a failed camera on its own.

What we develop

Six areas of one technology. Open an item for the technical detail.

Edge inference engine A multi-channel engine written in C++. Hardware decoding and NPU inference for many cameras on one device, with the CPU left free for rules and communication.
Input
RTSP and ONVIF, H.264 and H.265, any camera brand
Decoding
hardware video unit with 2D scaling
Inference
INT8-quantised models on the NPU, run asynchronously
Tracking
a track per object, ByteTrack
Channel budget
a camera is accepted only while both decoding and analysis capacity remain
Alarm transmission Our own protocol between the device and the monitoring station: two-way, authenticated and supervised.
Channel
two-way and authenticated, every event acknowledged
Supervision
heartbeat every minute, loss of a camera or of the panel link reported
Event
object class, rule, frames before and after, video clip
Delivery
queued and retried with back-off while the station is unreachable
Video search and forensics Searching recorded footage by plain-language description or by similarity to a chosen object. It runs locally, and originals are never modified.
Search
by description, using an image–text model, or by similarity to a selected object
Vehicles
make, model, colour, body type, plate and lettering
Cascade
motion, detection, object fingerprint, vision–language description; each stage runs only where the previous one found something
Integrity
SHA-256 of every file, activity log included in the report
Off-grid observation points A camera with on-board analytics on a mast, powered by solar and connected wirelessly, for places with no mains power and no fibre.
Power
PV with MPPT charging, LiFePO4 storage below the frost line, sized for December
Link
LTE or point-to-point radio
Structure
mast and mounts designed for image stability in wind
Rendering: a roadside mast with a solar panel, a camera and an antenna.
Viewing geometry Field of view and DORI distances computed from the lens and the sensor, before a camera goes up.
Method
DORI distances per IEC 62676-4
Data
optical parameters for more than 66,000 camera models
Output
coverage, blind spots and drawings in DXF
Model training and validation We train on our own footage and compare every model version on a fixed test set before it reaches a device.
Data
night infrared, rain, snow, birds right at the lens
Validation
the same test footage for every version, compared before release
Deployment
quantised for the NPU and shipped as versioned device images

How we build

Video stays on site

Analysis happens on the device. The station receives an event with a few frames, not a continuous stream: less bandwidth, and less personal data in circulation.

We work with what is already installed

RTSP and ONVIF from any camera maker, alarm panels, VMS platforms and monitoring stations. Our technology adds analytics to an existing system instead of replacing it.

Hardware and software in one pair of hands

We assemble, prepare and test devices ourselves. The engine, the models, the station protocol and the updates are our own, so we know what happens at every step.

DEFENSAR Technology

An engineering team from Pomerania, Poland. We work on video analysis from the video decoder to the protocol that carries an alarm to the station, together with security integrators, monitoring stations and infrastructure operators.

NCAGE code
9DBKHNATO Commercial and Government Entity code, assigned by the Polish Military Centre for Standardisation, Quality and Codification.
Location
Tczew, Pomerania, Poland

Contact

Tell us about the site: what it is, how many cameras it has, and which monitoring station you work with.

biuro@defensar.pl