Detect and track.
Dual-stream detection and multi-object tracking on the aircraft, so observations exist even when the link does not.
On the aircraft, Vektra is engineered to run beside the autopilot, not inside it. Detections, tracks and evidence are produced on board, shared with the flight stack over standard robotics interfaces, and held locally when the link is gone.
The heaviest decisions about an aerial model are made before it flies: which module it runs on, how much power and heat it can spend, and what it must do when the operator’s feed goes dark. Vektra builds the model layer and the onboard pipeline together.
Most public imagery is taken from head height. Aerial recognition is its own domain, and the dataset has to be built for it.
Objects shrink quickly as the aircraft climbs. Datasets span several survey altitudes, with scale augmentation in between.
Nadir, oblique and low-angle views. An object seen from overhead looks nothing like the same object seen from a doorway.
Paired visible and thermal imagery, with thermal-style augmentation of visible data where pairs are scarce.
Motion blur, haze, rain, low light and partial occlusion, represented in training and in validation.
Empty scenes and objects that resemble the target class, collected from the air on purpose.
A frozen aerial validation set, compared release by release on the target module.
A substation from an inspection drone: the whole frame, one transformer bay, one radiator bank. Altitude and angle decide what a class looks like, so the dataset holds every scale the aircraft flies.
Access lane, control building, three transformer bays and the switchyard beyond, from the inspection altitude.
One transformer with its radiator banks, conservator and bushings: the class an inspection model is trained to find.
Component scale: radiator fins, conservator tank and the cable trench beside them. Condition review starts here.
Dual-stream detection and multi-object tracking on the aircraft, so observations exist even when the link does not.
Larger second-stage models refine the class, re-associate tracks across views and rank candidates for review.
A person confirms or dismisses. Confirmation creates the evidence record, with its hash, time and position.
Most recognition failures are predictable. They cluster around scale, shift, clutter, concealment, sensor physics and the link to the operator. We treat each as a design requirement.
Read the field noteWide-area drone and satellite frames put most objects of interest under 32 pixels.
02 / DOMAIN SHIFTPublished benchmarks show detectors losing a large share of their accuracy when climate, terrain, season or sensor changes, with missed detections growing fastest.
03 / CLUTTER FALSE POSITIVESGrid-based detectors reason locally, so shadows, rubble and scrap routinely trigger false alarms.
04 / CAMOUFLAGE & CONCEALMENTConcealment is designed to defeat both eyes and models.
05 / THERMAL CONTRASTInfrared is strong against foliage and light fog, and weak when a target matches its background.
06 / DENIED LINKSJamming severs video and navigation links at the worst moment.
Export controls. Defense articles, technical data and defense services may be subject to the International Traffic in Arms Regulations (ITAR) and the Export Administration Regulations (EAR). We provide them only as authorized under U.S. export control laws, and we review end use and end user before any transaction. Do not send export-controlled technical data unless you are authorized to do so. Terms of use