Investigation
Bend Privacy Alliance / Axon patents series, Part 3 of 11 · September 2026
Inside Axon’s growing location and spatial-intelligence architecture
A police camera has a location.
At first, that sounds almost trivial. A body camera can record where it was. A patrol car can carry GPS. A license-plate read can be associated with a place and time. A map can put a dot where an officer or vehicle happens to be.
But that description increasingly misses what location does inside Axon’s expanding public-safety platform. Axon’s current Body 4 documentation distinguishes location embedded in recorded evidence from location sent into live operational systems such as Respond and Fusus. Fusus documentation describes maps that combine Body, Fleet and Air devices, users, CAD calls, floor plans and alerts. Other Axon documentation shows location being used to associate evidence with calls for service, reconstruct asset movement and route drones toward officers requesting help.
Axon’s patent portfolio goes further, describing architectures in which geography can help trigger camera recording, alter the frequency of location updates, preserve a camera’s movement when ordinary satellite positioning becomes unreliable, or predict which cameras may see a tracked person or vehicle next.
That is a different concept of location.
It is not simply a dot on a map.
It is increasingly a way for the system to decide what happens next.
The distinction matters in Bend because the City has spent years assembling pieces of the same broader platform. Bend contracted for Axon body cameras in 2021, Fleet cameras in 2022, Fusus in 2023, Axon Air-related services in 2024, and later a larger bundled Axon agreement. A City software inventory produced through Public Records Request 2026-257 also lists Axon Respond Plus as a Police Department production system, with a clipped description beginning that the product provides real-time updates on location. The inventory does not tell the public which location functions are enabled, whose locations are displayed, how frequently they are reported or how long those data are retained.
Those unanswered questions become more important once location is understood not as a single feature, but as an operating layer.
The first question: where is the camera?
Axon says Body 4 uses GNSS and optional Wi-Fi positioning. The same documentation distinguishes between location embedded in recorded evidence and location sent into live operational systems such as Axon Respond or Fusus.
That distinction is easy to overlook.
A location attached to an evidence file answers one question: where was the recording made?
A live operational location answers another: where is the device now?
Axon’s Respond documentation says Body cameras update location approximately every 10 seconds while recording and every 15 minutes while buffering. A separate Body 4 settings guide shows that administrators can configure “Location Updates during Buffering” and control when camera location is available on the Respond map.
That means a device can create operational location information while buffering even though Axon separately says location is attached to video only while recording.
Fleet systems make a similar distinction: the evidence record and the live operational map are not necessarily the same location dataset.
So even before getting to patents or predictive analytics, “camera location” already divides into multiple streams: location retained with evidence, location exposed on a live map, periodic buffering-state location and a device’s last known position.
Those streams do not necessarily have the same retention rules.
That is why one of the central findings of the underlying research is deceptively simple: a display window is not a retention period. A device disappearing from a map does not prove its location data were deleted. A system limiting a history query to one day does not prove the underlying history lasts only one day. A temporary map object expiring does not prove its audit record disappeared with it.
What happens when GPS stops working?
One Axon patent family takes the location problem a step further.
WO2024173733A1, “Selective location detection using image data,” addresses what can happen when a wearable or mobile recording device moves into a building, tunnel, urban canyon or other environment where ordinary satellite positioning becomes weak or unavailable.
The basic idea is not that the camera looks at a building and recognizes its street address.
Instead, the system can begin with a recent trusted location and then analyze how the camera itself moves.
The patent describes visual odometry, and potentially visual-inertial odometry, to estimate movement from changes in captured imagery and motion-sensor information. Buffered imagery may allow the system to reconstruct movement that occurred shortly before the device formally detected that its normal location signal had degraded.
Conceptually, the sequence is straightforward: a camera has a reliable absolute location; the location signal deteriorates; the device has continued capturing or buffering imagery; the system analyzes changes across those images; and estimated movement is applied to the last reliable location.
That can produce a continuing or reconstructed path for the recording device itself. It should not be confused with identifying or geolocating arbitrary people visible in the scene.
The accountability implications point in both directions. A more accurate record of where an officer moved inside a building could help reconstruct an encounter and test competing accounts of what happened. The same capability could also create location records in spaces where ordinary GPS would otherwise provide relatively little information.
The key provenance questions then become: Was a point measured directly by satellite? Estimated from Wi-Fi? Reconstructed from imagery? Was it interpolated? What was the accuracy? Was it current or merely a stale last-known location?
The research did not find current public Axon Body documentation confirming that this visual-odometry architecture has been commercialized in Body 4. The patent is evidence that Axon developed and sought protection for the architecture, not proof that Bend or any other particular agency currently uses it.
When location stops describing and starts triggering
A second Axon patent family illustrates an even larger shift.
US20250071509A1, “Using multiple geofences to initiate recording devices,” is a pending Axon application describing multiple geographic boundaries around a call for service.
The larger area functions as a safety or relevance zone. It can identify recording devices close enough to an incident to receive information about it. A smaller activation area can then determine when a device should begin recording.
The patent also describes variations in which location-reporting frequency can increase and the size of the geographic areas can change depending on how many recording devices are available or already recording.
That architecture changes the role of geography.
Location is no longer just information the system displays after something happened.
It can become an input to an automated rule: device enters area → system action follows.
The action may involve recording, communications or changes in how frequently location is reported.
It is important not to leap from that patent to a claim that Bend cameras automatically begin recording when officers cross a GPS boundary. The research did not find a current named Axon product feature publicly documenting the full nested-geofence recording architecture described in the pending application.
Axon does, however, already document commercially deployed geofencing for another purpose. Its Auto-Tagging geofencing guide says the system can compare recorded evidence with dispatch geography, using a configurable radius and the percentage of footage occurring inside that area to help associate evidence with a call for service. Axon documents a minimum configurable radius of 50 meters and permits exclusions such as hospitals or police headquarters.
That is a different function: location path + dispatch geography → evidence association, rather than location + boundary crossing → recording begins.
The difference is technical, but it is also a governance issue. The research identified several distinct mechanisms that can easily be blurred together under the word “geofence”: Axon Signal activation, Bluetooth-based Nearby BWC Activation, Auto-Tagging geofences, Fusus user-drawn geographic areas, Fusus Alert Regions and the patent architecture for nested recording geofences.
Each does something different. Each should have its own policy, configuration record and audit trail.
The map is becoming an operating environment
Location becomes more consequential still when individual devices are placed inside Fusus.
Bend approved Fusus in 2023. The City’s Connect Bend Privacy FAQ distinguishes registered cameras from integrated ones. Registration can place a camera on a map without continuous live access. Integration can allow conditional real-time streaming under the participating owner’s settings. The Connect Bend Data Share and License Agreement shows that integration may involve technical connection information and installation of Fusus equipment at participating properties.
That distinction matters because a camera map is one level of capability.
A live spatial operating environment is another.
Current Fusus map documentation says the system can show Body, Fleet and Air devices, users with GPS set to “On Shift,” active CAD calls, floor plans and alerts. Devices can appear as recording, buffering or inactive, and operators can center the map on a device’s last known location.
An alarm can move the operator to a device location and expose livestream or nearby-camera options. Fixed cameras and floor-plan context turn sensor placement into operational information, not merely inventory metadata.
A camera on its own records a scene.
A camera registered on a map becomes discoverable spatially.
A camera integrated into a live system can become remotely viewable.
A camera whose location is combined with an alarm or incident can be surfaced because the system judges it geographically relevant.
Once cameras are represented as spatially located and oriented nodes, their map position can support more than ordinary nearby-camera selection. The Fusus object-tracking patent family describes using that same spatial information to help follow a moving person or vehicle across the network: the current camera becomes the starting point, surrounding cameras are evaluated against geography and movement direction, and the system can identify which views are likely to matter next. That patent-described architecture goes beyond what a map merely shows; it uses the map as part of the decision about where an operator should look. US20250225666A1.
The surveillance capability emerges partly from the connection among those layers.
Police devices are not the only things that can appear on the map
fususOPS can expose user location separately from Body-camera location.
Axon’s On/Off Shift documentation says that when a user is on shift with location sharing enabled, their location is broadcast to dispatch and teammates on the map; when they go off shift, broadcasting stops. The same guide documents three GPS profiles: High, Moderate and Battery Saver, with different sampling and distance-filter settings.
That makes location an employee-privacy issue as well as a surveillance issue.
An agency may have a legitimate operational reason to know where personnel are during an emergency. But employee or BYOD location creates a different set of questions from evidence collected about a member of the public: Who can see it? Can the user pause it? Does an administrator have override authority? What happens when a shift ends? Does a last-known location remain visible? Are past locations retained? Are changes in the sharing setting audited?
Those questions cannot be answered by saying simply that Fusus “has GPS.”
From outdoor coordinates to floors and rooms
Ordinary GPS is also no longer the outer limit of Axon’s commercial spatial architecture.
Axon’s Body Mini indoor-location documentation describes a survey process in which specialists collect environmental signals including Wi-Fi, Bluetooth, cellular and barometric data and use them to create a calibrated map of a building. Axon says the result can provide floor-level and room- or zone-level location in surveyed areas.
That is materially different from a street-level map dot.
It introduces persistent spatial information about interior environments: floor plans, surveyed positions, calibration information and areas where location quality may be stronger or weaker.
Axon also acknowledges uncertainty. Its indoor-location guide says accuracy depends on building materials, access-point density and survey coverage, and identifies weak-signal areas where a device may report only a general venue or floor rather than a precise room.
A directly measured GNSS coordinate, a Wi-Fi-derived estimate, a calibrated indoor estimate, a stale last-known point and a predicted future route are not equivalent facts.
A defensible system should preserve the difference.
Then the subject begins to move
So far, most of the architecture has concerned the location of police devices, employees, fixed sensors or known assets.
The Fusus patent portfolio adds another category: the location and movement of people or vehicles being tracked.
The Fusus object-tracking patent family makes that spatial logic more concrete. US20250225666A1, a pending continuation in the same object-tracking lineage, describes a system that begins with a person or vehicle already visible in an active camera and then uses the surrounding camera network to help maintain the track.
In plain language, the system can combine where cameras are located, which direction they face, how the tracked object is moving, and where it is expected to go to identify cameras that may become useful next. The application describes determining direction of travel and a future route, while the broader family describes nearby or “next-up” cameras and predicted arrival timing based on factors including speed, camera location and field of view.
That changes the meaning of location again. The location of Camera B matters not simply because it can be drawn on a map, but because its position and orientation can make it a better candidate than Camera C for continuing the observation of someone leaving Camera A.
active camera + target movement + camera geography → likely useful next camera
The family also describes an AI-assisted reacquisition process for cases in which the tracked person or vehicle departs from the expected route or disappears from the predicted camera sequence. Nearby feeds can be analyzed for visible meta-attributes associated with the tracked object, and candidate matches can be presented to an operator for verification. That supports describing the patent concept as human-in-the-loop reacquisition, rather than assuming fully autonomous identity tracking.
For vehicles, patent-described attributes can include visible characteristics used to help distinguish one vehicle from another. For people, the specification discusses characteristics such as clothing, colors, hats, bags and other visible features. The important point is not that the patent proves every one of those functions is deployed in Fusus today. It shows how spatial prediction and appearance-based matching can be combined to help re-establish a track when geography alone is not enough.
Current public product evidence should remain separate from that patent disclosure. Axon’s commercial documentation does confirm that its vehicle-search systems already extract and search visual attributes. The current Fusus ALPR Search guide lists AI-detected color, make and type, plus “Additional Details” such as roof racks and bumper stickers. Fleet 3 documentation likewise says its ALPR system uses AI to establish vehicle color, make and type.
Axon’s broader Vehicle Intelligence materials describe the same idea in operational terms: a vehicle can be searched or distinguished not only by a plate, but by visible characteristics and other unique identifying features, including visible damage.
Those commercial capabilities should not be treated as proof that the patented reacquisition workflow is presently using those exact attributes to hand a target from one camera to the next. But they show that the two ingredients already exist within the broader ecosystem: spatial camera context and machine-readable appearance information.
Put together, the capability question becomes more consequential:
Where is the object moving, what does it look like, and which camera is most likely to see it next?
The distinction from device self-location remains important.
One asks: Where is the police camera?
The other asks: Where is the tracked person or vehicle going, and which camera should see it next?
Related Fusus research also identified WO2025128824A1, which describes video-based object tracking using isochrones, and US12348906B2, a granted Fusus patent covering dispatch augmentation with geolocated 360-degree static imagery.
Those patent disclosures should not be treated as proof that Bend has enabled every corresponding tracking feature. But they show why camera placement, orientation, routes and location histories matter inside an integrated network.
The network no longer needs to be understood only as hundreds of independent feeds.
It can be understood as a spatial graph.
A subject exits one view. The system has information about where the subject was moving. It knows where other cameras are. It can use geography, movement, camera orientation and, in patent-described reacquisition workflows, visual attributes to help decide which views may matter next.
That is the difference between having many cameras and having an architecture for navigating among them.
Location history is another form of surveillance history
Axon’s Track and Trace documentation adds another layer. It describes live and historical tracking of known mobile assets by VIN or ID, including current location, speed, coordinates, breadcrumb trails and a Trip History view with individual GPS pings.
The guide says users can review up to 24 hours of movement history at one time.
That sounds like a retention limit.
It is not.
The documentation limits the duration of a single history query, but it does not state how long the underlying GPS points are stored. A user can select a start and end date, and the system reports whether data are available for the chosen interval.
That distinction is especially important for location data because a collection of points acquires meaning through accumulation.
One observation may show where something was once.
A long history can reveal routes, routines, associations, stops and patterns.
Location data does not become sensitive only after someone runs a controversial search.
The history exists before the search.
And then location chooses a resource
One of the clearest examples of location becoming an operational command comes from Axon’s drone ecosystem.
Axon’s Watch Me drone-support documentation describes a workflow in which Fusus displays an officer’s Watch Me alert and livestream, DFR Command recommends the closest available drone, and the selected drone can fly autonomously to the officer’s location while a pilot retains the ability to take control.
The architecture is roughly: officer location → map context → nearest available resource → drone selection → autonomous route.
That is commercially documented spatial automation.
Bend has separately approved Axon Air-related procurement, and its later bundled Axon agreement placed more of the City’s Axon relationship into a common contractual structure. The reviewed records do not establish that Bend licenses or uses the Watch Me/DFR workflow described above.
That question should be answered with configuration and entitlement records, not inference from a product-family name.
But the commercial example illustrates the larger point.
Once location becomes an input to resource selection, accuracy and provenance are no longer abstract data-quality issues.
If a system decides that a particular camera is nearest, that an officer is on a particular floor, that a device is inside a boundary or that a drone should fly toward a particular coordinate, the public-interest question becomes: What location input caused that action?
Bend already has enough of the architecture to justify asking
None of this requires claiming that Bend has activated every Axon feature described in this article.
It has not been established.
That is precisely why the unanswered configuration questions matter.
The City’s public record already shows a substantial Axon footprint: body cameras, Fleet, Fusus, Axon Air-related services, Respond Plus in a City software inventory and a later multi-product bundle. The Connect Bend site also represents an expanding community-camera layer, while the City’s own privacy FAQ distinguishes simple registration from integrated, conditional live access.
The right question is therefore not whether Bend uses every patented or commercially available Axon feature.
There is no evidence for that.
The right question is whether the public knows what the platform Bend has already purchased is configured to do.
For location alone, local records should be able to answer whether Body-camera locations are transmitted while recording, buffering or both; how long operational location points are retained; whether Wi-Fi positioning is enabled; whether each point preserves its source and accuracy; whether Bend uses fususOPS employee location; whether last-known locations persist after active sharing stops; whether camera locations, floors and viewing directions are audited when changed; whether Auto-Tagging geofencing is enabled; whether Bend uses Track and Trace; whether any location rule can automatically activate a camera; and whether officer location can be used for drone selection or routing.
Most importantly, a later reviewer should be able to reconstruct not merely where something appeared on a map, but why the system believed it was there and what happened because of that location.
Appearance data complicates that picture further. A location point can say where a vehicle was observed; current Axon systems can also attach searchable characteristics such as make, color, body type, roof racks, bumper stickers and other visible features to that observation. A patent-described tracking system can then use movement, camera geography and visual attributes to help determine where that object may appear next or to reacquire it after a track is lost. The evidentiary categories therefore matter twice: a reviewer needs to know not only where a location came from, but also which observed or inferred characteristics caused the system to treat two observations as potentially involving the same object.
The problem is not the dot
A useful public surveillance inventory might once have been able to say that an agency owned body cameras, vehicle cameras and a map.
That is no longer enough.
A contemporary spatial system can contain multiple kinds of “location” at once: a satellite coordinate; a Wi-Fi-derived estimate; a mobile-user GPS point; a camera manually placed on a floor plan; a buffering device’s periodic location; a last-known location; an ALPR observation; a stored asset breadcrumb; an indoor room estimate; a user-drawn region; a dispatch geofence; a reconstructed device path; or a predicted area where a tracked person or vehicle may appear next.
Some are observed. Some are derived. Some are configured. Some are stale. Some are predictions.
Some may be used only for display.
Others can help determine what evidence is linked, which camera appears, which feed is shared, which resource is recommended or what device action follows.
A map that contains all of those things is not simply representing the world.
Parts of the system are acting on the representation.
That is why oversight has to reach deeper than whether a location feature exists. It has to ask what kind of location it is, how it was produced, how accurate it was, how long it survives, who can see it and what the system is permitted to do because of it.
When location becomes part of the logic that organizes evidence, surveillance and response, public oversight cannot stop at the map. It has to reach the rules that make the map act.