Solutions · Connected devices
For hardware makers: embedded movement AI, from one lab
Movelytics builds movement AI that runs inside a product: a mirror, a training machine, a kiosk. The model runs on the device's own embedded computer, with its own camera, without sending video to a server. We work on the hardware and the software together. Client hardware projects are under non-disclosure agreements, so none is named here.
Last updated:
Vertical jump · knee and ankle anglesIllustration. Authored pose sequence, not live inference.
01Your problem
A camera on a machine is not yet a product
Adding a camera to a mirror or a machine is the easy step. Making that device understand the person in front of it, in real time, on an embedded processor chosen for its price and its heat, for years, is a different project.
The constraints are harder than on a phone. The processor is fixed for the life of the product. There is no app store to push a fix. The device may sit in a gym with poor connectivity, or in a place where filming people and sending the video elsewhere is not acceptable.
02What becomes possible
What the technology makes possible
- Inference on the device's own computer
- The same approach as on a phone: the frame is analysed where it is captured. See on-device AI at Movelytics.
- A model sized for the processor
- We train our own pose model and run it with 8-bit inference, which is what modest processors execute fastest.
- Keypoints chosen for the machine
- A fixed camera and a known exercise allow a model and a logic tuned to that exact situation.
- No dependency on a connection
- The model needs no network to run, so the product keeps working when the network does not.
03Which route
A hardware plus software project with the lab
01
Scoping call
The device, the processor, the camera position, the movement, the constraints. We say what is feasible.
02
Proof of concept
A build that runs on your target hardware, with measured frame rates, not a slide.
03
Production
Optimisation for the final hardware, integration with your software, and maintenance.
We do take on hardware plus software products. Because these projects are under non-disclosure agreements, we do not name or describe them. What we can show publicly is how we measure: the benchmarks of our own model on named phones, and the research note on real-time pose on a low-end Android phone.
04Privacy
A camera in a shared space
A device with a camera in a gym, a clinic or an office raises an obvious question: where does the video go? With on-device inference the answer is short. It is analysed inside the device and it is not uploaded for inference.
That answer is an asset for your product as much as for ours. Our data commitments describe what Movelytics collects and what it never collects.
05FAQ
Questions from hardware makers
Does Movelytics build hardware products?
Movelytics builds hardware plus software products with clients. These projects are under non-disclosure agreements, so the company does not name or describe them publicly.
Can the model run on an embedded processor?
The lab trains its own pose model, designs it for modest processors and runs it with 8-bit inference. Whether a given processor reaches real time is measured in a proof of concept on the target hardware.
Does the device need to be online?
The model needs no network to run. A connection is only needed for what surrounds it, such as updates or usage reporting.
Connected devices
Tell us about the device.
Processor, camera, movement. We will say what is feasible.