Do AI-Tracking Webcams Help Local AI and VR? What the AI in OBSBOT and Insta360 Cameras Actually Does
Do webcams labeled “AI tracking" or “AI-powered" help with local AI or VR? Take an app like HIROMEIRO, which moves your whole body using a single camera. What does the camera’s “AI" actually do for it?
In this article I went through the official information on five AI-tracking cameras from OBSBOT and Insta360 and sorted out what the features marked “AI" really are. I also looked at how they differ from the AI that runs inside your computer (pose estimation and local LLMs), and at UVC, the common USB camera standard that connects the two. I did not use any of these cameras myself. This is a summary of publicly available information.
This reflects the state of things as of October 12, 2026.
- 1. Method: what I looked into and where the sources come from
- 2. What does a camera with “AI" actually do?
- 3. Does that AI run inside the camera or inside the computer?
- 4. The joint is UVC: what can an AI-tracking camera turn off?
- 5. How do you show camera video to a local LLM?
- 6. Discussion: where does AI help, by use case?
- 7. What I could not find out
- 8. Summary: does the AI in AI-tracking cameras help local AI and VR?
- 9. References
Method: what I looked into and where the sources come from
I looked into three things: what the “AI" in an AI-tracking camera concretely does, whether that processing runs inside the camera or inside the computer, and how to hand the camera’s video to AI running in the computer.
- Cameras covered: five models, OBSBOT Tiny 2, Tiny 2 Lite, Tail Air, Insta360 Link 2, and Link 2 Pro.
- Sources: I gave priority to the makers’ official pages and official manuals. Where an official page could not be read, I say each time that the item rests on secondary information such as a search-result summary.
- AI on the computer side: I read the official HIROMEIRO documentation, Google’s official MediaPipe page, and the official Ollama documentation.
My thinking on choosing a camera is in the first article below. My investigation of HIROMEIRO itself is in the second.
Which PC Webcam Should You Buy? Fixed, Wide-Angle, AI-Tracking, and Auto-Framing Types Compared
Full-Body Tracking with a Single Camera? What I Found Out About the Free Software HIROMEIRO
What does a camera with “AI" actually do?
The AI features the makers list are similar across all models: finding and following a person (AI tracking), adjusting the composition (auto-framing), and being operated by hand movements (gestures).
| Model | AI features the maker lists | How to turn tracking off (per the maker) |
|---|---|---|
| OBSBOT Tiny 2 | AI tracking with auto-zoom, gesture control, voice control | Voice command “Unlock Me", gesture, OBSBOT Center |
| OBSBOT Tiny 2 Lite | AI tracking with auto-zoom, tracking modes (upper body, close-up, headless, lower body, hand), zone tracking | Toggle with a gesture (OBSBOT Center also works) |
| OBSBOT Tail Air | AI tracking (people, animals, objects; animals are cats, dogs, and horses), Gesture Control 2.0, AI Director Grid | Release with a gesture, or double-tap in the app (official manual) |
| Insta360 Link 2 | AI tracking (one person or a group), gesture to start tracking | Toggle with the Touch Key or “AI Tracking" in Link Controller |
| Insta360 Link 2 Pro | AI tracking (people only), tracking-area settings, and more | Toggle with the Touch Button or “AI Tracking" in Link Controller |
All of these features amount to automatically moving the lens direction (pan and tilt) or the zoom. Pan and tilt mean the camera turning left and right and up and down. For an app like HIROMEIRO, which expects you to enter the field of view and the camera direction, this movement causes a mismatch with the entered values. I explained the reason in the earlier article.
Does that AI run inside the camera or inside the computer?
The specification section of the OBSBOT Tail Air’s official manual says “AI chip: Integrated AI chip". This reads as the AI processing being done on a chip built into the unit. For the other four models (Tiny 2, Tiny 2 Lite, Link 2, and Link 2 Pro), I could not confirm from official information whether the tracking runs inside the camera or inside the computer. The official page for the OBSBOT Tiny 2 says desk mode and similar functions are operated with buttons in the app, but that describes the entry point for the controls, not where the processing happens.
On the other hand, the AI that runs inside the computer is clearly documented. The official HIROMEIRO documentation says the processing that reads body movement runs in the tracking software. It runs on a Windows PC, Android, or a computer that can run Python. The video goes to a PC on the same network, and nothing describes sending it to the cloud. All it asks of the camera is video. MediaPipe’s pose estimation supports still images, video frames, and live input such as a camera. However, the official overview page does not say whether it runs on the CPU or the GPU, or how many frames per second it can process.
The joint is UVC: what can an AI-tracking camera turn off?
The common way to pass a camera’s video to AI in the computer is UVC, the shared standard for USB cameras. The UVC specification has items for zoom and for pan and tilt. I have not read the USB-IF specification itself this time, but I confirmed that the Linux kernel headers define control items for zoom, pan and tilt, and roll (rotation).
“AI tracking," on the other hand, is not a standard UVC item. It is a feature each maker added. The makers do not say whether it is driven through UVC controls or through a proprietary method.
The OBSBOT Tail Air has several ways to output video (Micro HDMI, USB-C, and NDI). The English version of the official manual (v1.0) says that when UVC output is enabled, NDI, recording, live streaming, HDMI, and RTSP cannot be used. When NDI is enabled, UVC, live streaming, and HDMI cannot be used. I could not check the Japanese version of the manual.
How do you show camera video to a local LLM?
Ollama officially explains how to pass an image to a model that can read images (a vision model). The SDK takes a file path, a URL, or bytes, and the REST API takes the image as a string in Base64 form. The official explanation uses gemma4 as its example. In addition, the official page lists qwen3-vl in sizes from 2B (1.9GB) to 235B (143GB).
The official documentation does not describe how to grab a single frame from a webcam feed and pass it in. Grabbing a frame with something like OpenCV and passing it in is not an official procedure, and I have not tried that combination on this blog either.
Discussion: where does AI help, by use case?
These are readings drawn from published facts. I have not confirmed them on real hardware.
| What you want to do | My reading | Reason |
|---|---|---|
| Move your whole body in HIROMEIRO | Tracking off, or a fixed camera | HIROMEIRO assumes you enter the field of view and the camera direction |
| Have the camera follow you for streaming or meetings | AI tracking helps | Following the person is the whole point |
| Show video to a local LLM | Tracking is probably unnecessary. It is enough if the video reaches you over UVC | Ollama’s image input handles how the image is passed in |
The “AI" in an AI-tracking camera is a function that moves the field of view and direction. When you use the camera on the local AI or VR side, whether you can turn tracking off looks more likely to be the deciding factor.
What I could not find out
- Whether the tracking runs in the camera or in the computer is not stated officially for the four models other than the Tail Air.
- For the Tail Air manual, I checked only the English version (v1.0).
- I could not read the body text of Insta360’s official product pages. The field of view and resolution are secondary information.
- I have not obtained how MediaPipe splits work between CPU and GPU, its processing speed, or the original text of the UVC specification.
- I have not used the five cameras on this blog, and I have not confirmed whether the field of view becomes fixed when tracking is turned off.
Summary: does the AI in AI-tracking cameras help local AI and VR?
The “AI" in an AI-tracking camera is a function that automatically moves the camera’s direction and zoom. The local AI side (HIROMEIRO’s pose estimation and Ollama’s image input) runs inside the computer and receives only video from the camera.
As a next step: if you already own an AI-tracking camera, check on the official page how to turn tracking off, then try whether the video still comes through over UVC in that fixed state. If you are about to buy one, make it a condition that the maker documents how to turn tracking off.
References
- OBSBOT Tiny 2 (official)
- OBSBOT Tiny 2 Lite (official)
- OBSBOT Tail Air (official store)
- OBSBOT Tail Air user manual (English v1.0)
- Insta360 Link 2 (official Japan store)
- Insta360 Link 2 AI Tracking (official FAQ)
- Insta360 Link 2 Pro AI Tracking (official FAQ)
- HIROMEIRO Quick Start (official, in Japanese)
- MediaPipe Pose Landmarker (official)
- Ollama Vision (official)
- qwen3-vl (Ollama library)
- USB Video Class 1.5 (USB-IF)