Latest-frame input
Threaded capture prioritizes the newest frame so stale imagery does not build up behind inference.
Technology // Engineering approach
Built from observations, not assumptions.
Reliable tracks begin with disciplined sensing. Lamassu processes camera data at the edge today and is engineering a fusion model that keeps sensor context visible as the system expands.
Compute // NVIDIA Jetson edge hardware
Sensing // Passive optical first
Edge pipeline
See. Decide. Follow.
The demonstrated prototype closes the loop between live camera input, computer vision, target selection, and physical optical motion.
Edge detection
A trained YOLO model detects drone candidates locally on GPU-enabled edge hardware.
Target persistence
Visual tracking maintains identity through a video sequence and exposes motion over time.
Closed-loop motion
Pixel error becomes bounded motor commands that move a calibrated pan-and-tilt optical mount.
Model evidence
Validation metrics, clearly labeled.
These results describe a held-out image validation set. They are engineering evidence, not a promise of performance in every field condition.
Single-class drone detector
Reported from the current trained model on held-out validation imagery. Real-world performance varies with range, target size, weather, lighting, background, camera quality, and dataset coverage.
94.8%Precision
95.1%Recall
96.9%mAP50
Network behavior
A tripwire mesh that asks for agreement.
The demonstrated subscriber treats one node as an observation and multiple timely nodes as stronger evidence, while monitoring whether sensors are healthy.
Require multiple observations
Two or more nodes reporting within a configured time window can produce a confirmed alert.
Reduce repeated alerts
Time-bucketed deduplication and cooldown behavior limit repeated messages from the same activity.
Know when a node is silent
Heartbeat tracking exposes nodes that have stopped reporting instead of quietly treating them as healthy.
Capability status
No blurred line between present and future.
The platform is early. We publish the maturity of each sensing and integration path so a technical conversation starts from the same facts.
| Capability | Status | Current boundary |
|---|---|---|
| Optical detection | Demonstrated | Live camera inference on Jetson edge hardware. |
| Visual tracking | Demonstrated | Video identity tracking and closed-loop optical follow. |
| Node corroboration | Demonstrated | Subscriber logic exists; the full node publisher path is still being completed. |
| Remote ID / RF | In development | Receive-only bench tooling; not yet integrated into the track pipeline. |
| Multi-camera 3D fusion | In development | Bearing-only triangulation architecture is designed but not yet operational. |
| Radar & acoustic inputs | Roadmap | Planned sensor adapters; no current operational integration. |
| Operator UI & C2 output | Roadmap | Planned structured track output; not currently available or certified. |
Technical diligence welcome
Bring the hard questions first.
We can discuss the current prototype, sensing constraints, validation boundaries, and the architecture required for your environment.
Start a technical conversation