LamassuSystems

01 · Detection & tracking

Prototype demonstrated

See what’s flying.

A passive, camera-first sensor node that finds small drones in live video, holds a track on them, and physically follows them with a pan-tilt mount — all on edge hardware, with nothing sent to the cloud.

How it works Request a demo
EO-01 · Sensor view Edge · Jetson Orin Nano

Illustrative simulation · Not live sensor footage

01 Pipeline

From pixels to a held track.

Four stages run on the node in a closed loop, and a fifth ties nodes together. Every new frame updates what the sensor sees — and where it points next.

  1. Stage 01

    Capture

    An electro-optical camera streams video. The node always works on the newest frame, so delay never piles up.

  2. Stage 02

    Detect

    A single-class YOLO model, trained to find drones, runs on the GPU and proposes boxes with confidence scores.

  3. Stage 03

    Track

    A multi-object tracker (BoT-SORT, or the lighter ByteTrack) links detections across frames, so each drone keeps one ID through brief misses.

  4. Stage 04

    Follow

    A closed-loop controller turns pixel error into motion on a two-axis stepper pan-tilt mount, within calibrated limits.

  5. Stage 05

    Corroborate

    Nodes report detections over MQTT, and a subscriber confirms a target only when independent nodes agree. Logic built; field tests next.

02 Validation

Measured — and honest about it.

Our current detector, measured on images it never trained on. Strong numbers, clear definitions, and the limits we’re still working on — all in one place.

Detector validation · Current model

Demonstrated
94.8 percent Precision
95.1 percent Recall
96.9 percent mAP@50

How we measured

Model
YOLOv8n trained for one class — drones — at 640 px input.
Data
Validation split of a ~22,000-image public drone dataset, held out from training.
Metric definitions
Precision
Of the boxes the model drew, the share that were really drones.
Recall
Of the drones in the validation images, the share the model found.
mAP@50
Average precision when a box counts as correct if it overlaps the true drone by at least 50%.

Engineering validation on still images — not a guarantee of field performance in every condition.

Known limits we’re working on
  • Confidence drops on distant, very small drones.
  • Haze, cloud, and low-contrast backgrounds make detection harder.
  • Dense swarms are not yet tracked reliably — training data is mostly single drones.
  • Live field accuracy still has to be measured at pilot sites.

03 The prototype

What’s on the bench today.

The current node is built from proven, off-the-shelf parts so we can iterate quickly on the parts that matter: the model, the tracker, and the control loop.

Compute
NVIDIA Jetson Orin NanoEmbedded GPU module running inference on the node.
Sensor
Electro-optical cameraPassive sensing — the node watches without emitting.
Model
Single-class YOLO detectorTrained to find one thing well: drones.
Tracking
BoT-SORT multi-object trackerPersistent IDs across frames; ByteTrack as a lighter option.
Actuation
Two-axis stepper pan-tiltClosed-loop follow inside calibrated travel limits.
Messaging
MQTT corroboration serviceSubscriber logic for detections, heartbeats, and alerts.

04 Network

One node sees. Two nodes agree.

Any single sensor can be fooled. Our corroboration logic confirms a target only when independent nodes report it within the same window. The rules below are built; field testing across multiple nodes is next.

Three sensor nodes corroborating one drone Nodes 01 and 02 both see the same drone inside their overlapping coverage, so the subscriber confirms the track. Node 03 does not see it. TRK 07 · 2 NODES NODE 01 · DETECT NODE 02 · DETECT NODE 03 · CLEAR SUBSCRIBER Track confirmed 2 NODES · 10 S SIGHT LINE MQTT MESSAGE
  • 2+
    Nodes to confirm

    A target is confirmed only when two or more nodes report it.

  • 10 s
    Corroboration window

    Reports have to land within ten seconds of each other.

  • 30 s
    Alert cooldown

    One event raises one alert, not a flood of repeats.

  • 60 s
    Heartbeat timeout

    A node that goes quiet for a minute is marked offline.

05 Status

What works today, and what’s next.

Every capability carries its real stage. We’d rather show you an honest table than a polished promise.

Detection capability status
Capability Stage Notes
Edge drone detection Demonstrated Single-class model running on a Jetson Orin Nano.
Multi-frame tracking Demonstrated BoT-SORT keeps persistent IDs across frames.
Pan-tilt follow Demonstrated Two-axis stepper mount with calibrated limits.
Multi-node corroboration In progress Subscriber applying the two-node, ten-second rule is built; multi-node field tests are next.
Remote ID reception Bench tooling Receive-only tools for broadcast Remote ID.
Long-range small targets In R&D Distant, very small drones remain a known weak point.
Swarm tracking Planned Dense groups are not yet tracked reliably.
Shared node with charging Planned One ground node that senses and charges.

06 Where it fits

Built for sites that can’t watch the sky all day.

The applications we’re designing for. Each one starts with a pilot site and a conversation.

Critical infrastructure

Substations, water, and energy sites that need to know what’s overhead.

Industrial campuses

Large footprints with drone traffic — their own fleets and everyone else’s.

Events & venues

Temporary airspace awareness for crowds, stadiums, and outdoor events.

Drone operators

Know what else is in the air before your own fleet launches.

Detection only. No weapons.

Lamassu builds sensing and awareness. Our systems detect, track, and alert — they do not intercept, jam, or engage aircraft. Any response stays with the people and authorities responsible for the site.

07 Next step

Put a sensor on your skyline.

We’re looking for pilot sites and partners to put the detection node through real-world conditions — and to shape what it becomes.

Request a demo See autonomous charging