Peregrine Labs
Edge AI for safety and security
We put perception on cameras that have to work where the network drops and nobody is watching the screen.
- Fixed installationsSites where the link goes down and the camera has to carry on.
- AirBeyond the range of reliable connectivity.
- GroundWhatever compute the vehicle happens to carry.
The gap
Most footage is reviewed long after it mattered
Control rooms and field teams hold more video than anyone can watch. The frame that mattered turns up later, in an export.
Understood at the camera
INSIDE A SECOND
Recorded now, read later
HOURS TO DAYS
Protect life
A warning reaches a crew while the situation can still be changed.
Protect critical infrastructure
Perimeters, depots and substations are watched by cameras that record far more than they report.
Keep operating under pressure
Fewer false alarms mean the ones that do arrive get read.
Where we work
Fixed, airborne and on the move
We tune our perception foundation to the platform it runs on and the event it has to catch.
FIXED INSTALLATIONS
Cameras that stay where they are
Perimeters, public spaces, depots. The camera sends the event near real-time.
- Perimeter and zone logic
- Crowd and flow
- Anomaly alerts
AIR
Drones and airborne systems
Reconnaissance, emergency response, inspection.
- Detection and tracking
- Change between passes
- Overlays for the crew
GROUND
Vehicles and working sites
Patrol and response vehicles, yards, off-highway machines. Perception on the vehicle's own compute.
- Person and vehicle proximity
- Restricted zones
- Site safety events
On the device
From pixels to a decision someone can act on
Detection on its own produces alerts nobody trusts. Security work needs position, movement, zone and timing before an alert is worth sending to a person.
ALERT
- CLASSIFICATIONperson on foot
- WHEREon the marked crossing
- INTERACTIONlorry on a converging path
- GAP15 m, closing
Clip attached. Faces blurred on the device.
-
Objects and people
Detected, classified and tracked across frames.
-
Movement and behaviour
Direction, speed, dwell and interaction.
-
Zones and relations
Position against a boundary or an asset.
-
Change over time
What differs between one pass and the next.
-
Anonymisation
Faces and plates blurred before storage.
-
Event records
Structured, time-stamped, searchable.
Why the edge
Built for sites where the cloud is not an option
Safety and security systems run with hard limits on connectivity, response time and what is allowed to leave the site.
At the camera
DETECTING
- Sees the scene and classifies it
- Applies the zone and event logic
- Raises the alert
Elsewhere
SYNCED OFFLINEDashboards, archives and reporting catch up when the connection returns.
Perception runs locally
A dropped link delays the reporting but does not stop the detection.
The response happens where the event is
No round trip to a server before a warning reaches the crew or the signage.
Data stays on site
Processing can run entirely on infrastructure you control, including sites with no connection at all.
- 95%
less bandwidth than streaming video, in selected deployments.
- 90%+
fewer false alerts, in selected deployments.
Evidence
In operation outside the lab
Multiple deployments outside road transport, with the organisations that run them.
AIR
The scene, before the crew arrives
Live drone footage with detection overlays, streamed into fire brigade vehicles on their way to an emergency.
WITHGaia-X 4 AMS consortium, local fire services
Read the case
FIXED INSTALLATIONS
A public space that warns only when the risk is real
Sensors and adaptive signage watch a school zone, and warn when the situation calls for it. Drivers stop tuning the signs out.
WITHT-Systems, Landshut University of Applied Sciences, City of Landshut, backed by the BMDV
Read the case
GROUND · DEMONSTRATION
Perception on a body-worn camera
Anonymisation, detection and pose tracking running on the camera itself, inside the power a worn device carries.
WITHBuilt by Peregrine Labs as a working demonstration
Discuss a body-worn buildHow we work
What a Labs project looks like
Three stages, each with a fixed output and you decide whether to start the next one.
-
01 Scope
Decide what has to be understood
We find out whether perception on your hardware is feasible.
Targets agreed before anything is built. OUTPUTFeasibility assessment and deployment plan
-
02 Build
Adapt the perception to your platform
Models are tuned to your objects, your scenes and your chip.
Cut until it fits the power and thermal budget it has to run in. OUTPUTWorking prototype on your platform
-
03 Validate
Prove it where it has to run
We run it in the field until your operators trust what it sends them.
False alerts per shift, across a field trial. OUTPUTValidated system and a route to scale it
Integration
Plugs into your control room
We add perception to your cameras and send the result to your operators.
INPUT
Continuous video
ON THE DEVICE
Peregrine Vision
WHAT LEAVES
Structured events
- Person in a restricted zone
- Vehicle entering the lane
- Gate state changed
OUTPUT
Your system
- VMS
- Mission control
- Dispatch
- GIS
- SCADA
- Your dashboard
Two frames in forty were worth sending. The rest never left the camera.
Hardware-agnostic
Embedded GPU, CPU, NPU or ASIC.
Updates over the air
Reach devices in the field.
Modular event logic
Change classes, zones and triggers.
APIs and SDK
Firmware, REST or webhooks.
Silicon we deploy on today. If your chipset is not here, ask us.
Privacy and sovereignty
Anonymised before it is stored
Faces and licence plates are blurred. All footage is encrypted.
- No facial recognition
Nobody is identified. Faces are removed.
- Deployable under GDPR and the EU AI Act
Written into the design rather than added at the end.
- Your infrastructure, your country
Processing can run on hardware you own, on sites with no connection.
Who we build with
Partners who operate outside the lab
Operators, integrators and research institutes we have delivered with.
DELIVERY AND RESEARCH
PROGRAMMES AND FUNDING
Research
The people we work through the hard problems with
Joint research with three German universities, on perception, data and learning from real scenes.
In early 2023, we hired the first Ph.D. student conducting cutting-edge research at the intersection of AI and sustainability as part of a joint program between HPI and Peregrine.
Professor Ralf HerbrichHasso Plattner InstituteWe have successfully conducted joint research on Self-Supervised Contrastive Learning in Traffic Scenes, experimenting with new approaches for Instance Segmentation using unique data provided by Peregrine.
Professor Cyrill StachnissUniversity of BonnWe worked with Peregrine and Seagate to experiment with real-time data logging and Peregrine's smart data campaigning system for real-time data annotation using our self-driving vehicle “MadeInGermany”.
Professor Daniel GöhringFreie Universität Berlin