Investigation
Bend Privacy Alliance / Axon patents series, Part 2 of 11 · September 2026
Axon’s patents describe camera systems that can change what is preserved before recording, generate selected detail, infer what a wearer could see, and even construct video intended to simulate human perception.
Evidence boundary: A patent establishes that Axon claimed or disclosed a capability. It does not establish that Axon currently sells that capability, that a current camera implements it, or that Bend Police or another agency has activated it. This article distinguishes patent evidence from current product documentation throughout.
A police-camera video seems like one of the simplest records to understand.
A lens was pointed at a scene. Light reached a sensor. The camera encoded what it captured. Later, someone pressed play.
That description is becoming incomplete.
What if the seconds before an officer formally begins recording were captured at one quality, then automatically switched to another quality because the camera detected a change in circumstances?
What if one image inside a continuous-looking video was produced through a different processing path because a particular object or region had been selected for greater detail?
What if the video displayed a boundary showing what software calculated the officer could see?
And what if the recording being reviewed was not any camera’s original recording at all, but a new video generated from multiple sources to approximate human perception?
Those possibilities appear across a set of Axon camera and computational-imaging patents reviewed by Bend Privacy Alliance. Together, the patents describe a progression from ordinary image capture toward something more complicated: evidence whose content or presentation can depend on automated state changes, computational processing, sensor-derived inference, or reconstruction.
The important question is no longer simply whether a video is authentic. It is what kind of video it is.
That distinction matters especially in policing, where camera footage can influence charging decisions, use-of-force reviews, internal investigations, civil litigation, public understanding, and testimony about what an officer perceived.
“Video evidence” is becoming too broad a category
It is tempting to divide evidence into two categories: real and fake.
The patents examined here suggest that distinction is inadequate.
There can be an untouched source recording. There can be source video captured first at one quality and then another. There can be an image generated from a selected portion of high-resolution sensor data. There can be original footage carrying a computationally derived overlay. Multiple recordings can be aligned or combined. And a new video can be generated to represent something no single camera directly recorded.
These categories do not necessarily describe current Axon products. They are an oversight framework drawn from the camera capabilities described across the patent portfolio.
Each step changes what a reviewer needs to know. With ordinary source video, the basic questions concern the camera, timestamp, recording state, chain of custody, and whether the file has been altered. With computationally derived evidence, additional questions appear: What data went into the transformation? What triggered it? What software performed it? Were assumptions applied? Does an untouched original still exist? Can another person reproduce the result?
Before anyone presses record
Pre-event buffering itself is not hypothetical. Axon’s current Body 4 documentation says Ready mode captures a configurable pre-event video buffer and saves the buffered segment when Event recording begins. Axon’s Fleet 3 documentation likewise identifies Buffering as its default pre-event mode.
A newer patent family goes significantly further.
Axon’s application US20260122348A1, “Multiple buffering modes for video recording devices”, describes a system that operates first in one buffering mode, switches to a second buffering mode with higher video quality after one condition occurs, and enters recording mode in response to another condition. The independent claim makes conditional pre-event quality switching part of the claimed concept.
Dependent claims describe circumstances that can affect state changes, including decreased movement speed, determining that the user is exiting a vehicle, the state of another nearby recording device, acceptance of a task, battery or power conditions, location or proximity, and environmental sound level.
The evidentiary consequence is that pre-event video may look continuous when played back even though the system did not necessarily treat every second the same way while capturing it.
Imagine a camera buffering relatively low-quality video during routine activity. Something changes. Before the wearer formally enters Event mode, the system recognizes a condition and begins buffering at higher quality. Later, the resulting segments become part of an evidentiary recording.
The event file may tell a visually continuous story. Its acquisition history may be more complicated.
What existed in the buffer before formal recording, at what quality, and why?
For evidentiary review, it would be useful to know when each mode changed, the reason for the transition, the resolution or other quality characteristics associated with each segment, and whether any buffered material was overwritten before preservation.
There is an equally important product boundary. Current Axon documentation establishes pre-event buffering. The reviewed product materials do not establish that current Axon cameras implement the patent’s adaptive switching between different pre-event quality levels.
A frame inside the video may not be like the others
Another Axon application takes the problem from changing recording states to changing how a particular image is produced.
US20260162220A1, “Detail Capture During Video,” describes a recording device that identifies a detail subject and boundary and generates a detail frame through a distinct sensor-and-processing pathway during ongoing video capture.
The independent claim does not require automated object recognition. That appears in a dependent claim as one possible way to identify the detail subject. Other dependent implementations include binning and cropping.
Most importantly for evidence provenance, dependent claim 4 expressly states that the detail frame can be included in the video recording.
That means a continuous-looking evidentiary video could potentially contain an image produced through a different processing pathway from ordinary frames immediately before and after it.
That is not necessarily improper. A higher-detail frame could preserve information ordinary video might otherwise miss. But it changes what the record means.
Was the subject selected by the wearer? By software? Did the camera crop a region from a larger sensor image? Was the full source frame retained? What processing parameters were used? Can the special frame be distinguished from ordinary frames during playback?
Axon Body 4 provides a useful comparison. Its current High Resolution Photo Capture documentation says a 1440p photograph can be taken while the camera is recording or buffering, but that the photo is stored as an independent evidence file and is not linked to the recording session. That current feature should not be casually equated with the patent’s embedded-detail-frame embodiment.
What did the officer supposedly see?
The next patent moves from image enhancement into interpretation.
Axon’s granted US11632539B2, “Systems and methods for indicating a field of view,” concerns indicating a person’s field of view within video captured by a wearable-camera system. The patent identifies a February 14, 2017 priority date and lists Axon Enterprise as assignee.
The granted independent claim describes a camera field of capture wider than the person’s field of view, sensor data corresponding to head orientation, server-side determination of the person’s field of view using that sensor data and limits of the person’s vision, and display of the inferred field of view relative to the wider recorded scene.
One distinction deserves emphasis: the main claim does not require the system to analyze the wearer’s face. It can estimate where the person was looking from sensor data about the direction and angle of the person’s head. A separate, narrower claim adds another option: the software can examine visible facial features in the recorded video — such as the eyes, ears, nose, jawline, or chin — and use their position to help estimate which direction the person’s face was pointed. Other claims describe showing the resulting estimate on the video with a boundary, or visually darkening, blurring, or obscuring portions of the camera image that the system calculates were outside the person’s field of view.
The significant capability is therefore not facial recognition or identifying who the wearer is. It is using sensor data — and, in one version, the visible position of facial features — to make a computational estimate of what direction the person was looking and what part of the scene they could supposedly see.
The camera records the physical scene within its own field of capture. The system then uses other data and assumptions about human vision to calculate which portion of that scene the wearer allegedly could perceive.
Those are not the same thing.
A recorded pixel can establish that something was inside the camera’s field of capture. It does not, by itself, establish that the wearer saw it. An inferred field-of-view overlay attempts to answer that second question computationally.
The patent itself explains why this can matter, including in reviewing what a person could see before a weapon discharge or use of force.
What head-orientation measurement was used? How precise was it? What range of human vision did the system assume? Was peripheral vision represented? Did the model change according to the individual or circumstances? Which software version generated it? Can a reviewer see the unmodified footage?
Current Body 4 product materials document a 160-degree camera field of view and an optional Flex POV accessory for additional perspectives. The reviewed current materials do not establish commercial deployment of the patented inferred-field-of-view overlay.
When evidence becomes a simulation
One patent family moves farther still.
Axon’s PCT publication WO2023244829A1, “Generating video data for simulating human perception,” describes systems that process video from multiple recording devices associated with a person at an incident and generate simulated video intended to represent aspects of human perception.
The disclosure discusses recordings with differing viewpoints and camera characteristics, and describes processing that can involve head orientation, gaze or eye-tracking information, spatial or metadata-based alignment, lighting adaptation, field-of-view changes, blur, stabilization, object recognition or machine-learning techniques.
The claim language is now independently verifiable. Claim 1 expressly recites receiving video from a first camera and a second camera worn by the user and combining those recordings to provide simulated video data that indicates the user’s perception at the incident. Dependent claim 6 further states that the represented perception can include at least one of head orientation, gaze, or lighting adaptation of an eye. Other disclosed processing details remain specification-level unless tied to a particular claim.
The last object in that chain is fundamentally different from either original recording. It is not merely another camera angle. It is a computationally generated representation.
Before relying on the image, a reviewer should know whether it is recorded evidence or a computational hypothesis about perception.
If such a system were ever used in an evidentiary workflow, every underlying source recording should remain available. Processing steps and software versions should be documented. Additional sensor information used in the reconstruction should remain linked to the output. And the derived video should carry a separate identity from the source footage.
The reviewed product materials did not establish that Axon currently offers this simulated-perception capability as a deployed evidence feature.
Not every camera patent is about reconstructing perception
The broader portfolio also helps put these examples in proportion. Bend Privacy Alliance’s research resolved the Camera Systems & Computational Imaging category into seven canonical families.
Axon’s dual-mode camera / quasi-bandpass filter family, for example, concerns visible and near-infrared imaging. Axon’s current Fleet 3 documentation independently establishes that the interior camera has an ambient-light sensor and engages infrared illumination in low light. That functional overlap does not prove that Fleet 3 practices the patent’s specific optical claims.
Another family, Flexible recording systems, describes primary and auxiliary recording devices and multiple possible camera configurations. Current Body 4 materials advertise an optional Flex POV module that plugs into the camera and records from additional angles. Again, that is functional overlap, not a claim-by-claim patent-to-product mapping.
The distinction prevents the patent research from becoming a catalogue of worst-case implications. Some patents describe capture hardware. Some describe processing. Some describe inference. And some contemplate creating a new evidentiary representation.
What Axon documents today—and what it does not
The product comparison gives us both positive and negative evidence.
Current Axon materials independently document body-worn recording, pre-event buffering, multiple perspectives through Body 4 and Flex POV, Fleet 3 interior infrared illumination, and high-resolution still-photo capture while Body 4 is recording or buffering.
The reviewed product documentation does not establish adaptive multi-quality pre-event buffering, patent-style detail frames embedded in ordinary video, post-event inferred field-of-view overlays, or simulated human-perception video.
Patents are easy to misuse in both directions. A patent is not proof of a current product. But an uncommercialized patent is not automatically irrelevant to oversight. Patents can document technical research, design directions, and capabilities a company sought legal protection for. What they cannot tell us without additional evidence is whether a particular customer has access to those capabilities today.
Why this matters in Bend even without evidence that Bend uses these features
Nothing reviewed for this article establishes that Bend Police is using adaptive multi-quality buffering, Detail Capture frames, inferred field-of-view overlays, or simulated human-perception video.
The local relevance comes from a different fact: Bend’s relationship with Axon has not remained frozen at the camera system the City originally approved.
In 2021, Bend approved an Axon body-camera agreement. In 2022, the City approved Fleet cameras. In 2023, Bend approved a Fusus agreement; the City-hosted Fusus terms are part of that procurement record. In 2024, Bend expanded Axon Air and later authorized a broader Axon Officer Safety Plan bundle. A related Third Amendment and consolidated quote placed previously separate products into one broader contractual relationship.
That history does not prove that every future Axon feature automatically arrives in Bend. It shows why product evolution matters to public oversight.
A five-year camera contract is not necessarily a five-year freeze on what software can do. Hardware changes. Firmware changes. Cloud services change. Licenses change. New analytic and AI functions can be created. Existing systems can become integrated with systems that did not exist when the original acquisition was discussed.
What material changes would require the public to be told before tomorrow’s camera becomes capable of producing a materially different kind of evidence?
What an evidence record should be able to answer
The patent portfolio suggests a practical test that is more useful than simply asking whether a video has been “edited.”
- Is this untouched source video?
- Did any portion originate in a pre-event buffer?
- Was the buffered material captured under more than one quality or processing mode?
- What caused any mode change?
- Does the sequence contain a specially generated detail frame?
- Was any object or region selected for different processing?
- Is any field-of-view boundary displayed on the video?
- Is that boundary measured directly or computationally inferred?
- What sensors, parameters and assumptions produced the inference?
- Has footage from multiple cameras been aligned or composited?
- Is any portion intended to simulate the wearer’s perception?
- Are every underlying source recording and relevant sensor record still retained?
- Is the software or model version that produced the derived output preserved?
- Are transformation events contained in the audit trail?
- Can an independent reviewer reproduce the result?
- Can the source evidence be viewed without the derived layer?
Those questions do not assume the transformation is deceptive. They recognize that different evidentiary objects require different provenance.
A photograph can be authentic and still be a crop. A video can be genuine and still contain a computational overlay. A derived reconstruction can accurately follow its algorithm and still depend on debatable assumptions.
Provenance is what allows the viewer to distinguish those questions.
The camera and the computation
Body-camera evidence once seemed conceptually simple because the chain was easy to describe:
Axon’s patent portfolio describes something more complicated.
Software can potentially affect what is preserved before formal recording begins. A selected part of a sensor image can be processed differently to create a higher-detail frame. Sensor information and assumptions about human vision can be used to calculate what a wearer supposedly could see. Multiple recordings and other information can be processed into a new video intended to simulate perception.
Some of those capabilities remain patent disclosures rather than documented commercial features.
That boundary is important, but so is the distinction on the other side.
If technologies like these ever move from patent records into ordinary evidentiary workflows, “the video” will no longer be an adequate description of what a reviewer has been shown.
The integrity of the record will depend not only on preserving the file. It will depend on preserving the difference between observation and inference, source and derivative, sensor data and software output.
What came from the camera, and what came from the computation?