Field Notes

concept

Visual positioning systems

StatusWorkingUpdatedStands on2 sourcesSupportsnotes

A visual positioning system determines where a camera is, and which way it is pointing, by matching what the camera currently sees against a previously built model of that place. It answers the same question as satellite positioning by an entirely different route: GNSS derives a location from timing signals broadcast from orbit, while a VPS derives it from recognizing the ground itself.

The category matters because it inverts a familiar privacy intuition. Satellite positioning is a receive-only operation, so a phone can know where it is without telling anyone. A VPS instead requires that someone has already built a detailed visual model of the place, which makes the reference map — not the fix — the asset, and makes whoever assembled that map the party with leverage.

How it works

The system extracts stable visual features from a live image — building edges, structural geometry, terrain patterns, the corner of a plinth — and searches for the same features in a reference model of the area. A successful match yields a full 6DoF pose: three coordinates of position and three of orientation.

That second half is what satellite positioning cannot supply. GPS can place a person within a radius of a statue; it cannot say whether they are in front of it or behind it, three metres away or thirty, or which way they are facing. Anchoring a virtual object convincingly to a real doorway requires the orientation as much as the position, which is why the technique grew up inside augmented reality before anyone marketed it elsewhere.

Reference models are built two ways, usually together: classical structure-from-motion reconstruction, and learned representations that encode a place implicitly in the weights of a small neural network rather than in an explicit point cloud.1

What it can and cannot do

Accuracy claims are conditional, and the conditions do the work. Centimetre-level localization is reported for well-mapped locations under good conditions; the same system in an unmapped or visually ambiguous place returns nothing useful at all.

Condition Behaviour
Densely scanned landmark, similar lighting Best case; centimetre-level pose
Mapped from one side only Degrades or fails from unmapped angles
Never mapped No fix; there is nothing to match against
Dense tree canopy, heavy occlusion Features obscured; unreliable
Featureless terrain such as open ocean or desert Little to match; unreliable
Scene materially changed since mapping Match quality falls as the model goes stale

The genuine privacy exposure is upstream, at map construction. Building the reference model requires imagery of real places, and where that imagery is crowdsourced it may include interiors, private property, and bystanders who were never asked. The exposure attaches to whoever contributed scans, and to whatever was in frame, rather than to people the finished system later positions.

Why the defense interest

Militaries want visual positioning because satellite positioning has become a dependency an adversary can attack. Jamming denies the signal, spoofing falsifies it, and dense urban geometry degrades it through multipath error without anyone attacking anything. A camera-based fix has no signal to jam, which makes it attractive as a redundant layer rather than a replacement.

The capability being procured is therefore orientation and coordination: letting a drone, a ground vehicle, and a dismounted operator share one coordinate frame when the satellites are unavailable. That is a genuine military capability and a real reason to scrutinize who supplies it. It is not the same thing as a targeting system, and treating a positioning layer as one misdescribes both what is sold and what to worry about.

Who builds them

No party owns the category, which is what makes this a concept rather than a note about a product. Niantic Spatial built its reference model from pedestrian scans contributed through games; Google’s Live View and geospatial APIs work from Street View imagery; Vantor’s Raptor localizes aircraft against satellite-derived 3D terrain rather than ground-level imagery. The differences that matter between them are the provenance, perspective, and coverage of the reference map, not the matching technique, which is broadly shared.

That is also the competitive logic: the map is expensive and the algorithm is not, so an organization that acquires ground-level coverage cheaply holds an advantage that is difficult to attack directly.

  1. Building a large geospatial model to achieve spatial intelligence, Niantic, 2024-11-12. ↩

Built on 2 sources (2 external).

Working out connections…