This OAK-1 and Raspberry Pi-Powered Device Forces You to Perform Push-Ups to Unlock Your Desktop
Driven by a pose estimation network and a classifier, this smart camera setup forces you to work out every hour — or give up altogether.
ClearML developer Victor Sonck has come up with a novel way to fight the work-from-home weight-gain: a machine learning system that won't unlock his workstation until a certain number of push-ups have been performed.
"I'm getting fat, so I can't be bothered to go to the gym or do anything productive really," Sonck explains. "So, I'm going to make a machine learning model. The plan is to lock my computer every hour or so then have a camera active somewhere, detect my actions, see that I did five push-ups, then unlock my computer again. Simple, right?"
The project is split across three key pieces of hardware. The first is Sonck's workstation, which only needs changes in software to lock on a schedule then wait for a signal from the machine learning system once the push-ups have been completed. The second is a Raspberry Pi single-board computer, which performs inference on the incoming video stream to detect the push-ups.
The final piece of the puzzle is the camera which captures the video: An OAK-1 depth-sensing camera, which offers both a color video stream and additional depth data missing from a standard webcam while offering an on-board accelerator for machine learning workloads — taking some of the strain off the Raspberry Pi without bogging down the target workstation.
"I [could] just attach any old webcam to a Raspberry Pi and have it run both the BlazePose model and the smaller classification model," Sonck explains. "But the Raspberry Pi isn't that powerful and I think it might catch on fire if I try to run both models. The OAK-1 is basically a 4K camera, so the imagery will be near-perfect, and then in the back it has an Intel Myriad X accelerator chip […] so it will be able to run BlazePose at, like, 15 frames a second which is more than enough for what we want to do."
There's a final trick to the system, too, which prevents any cheating. "'But,' I hear you say, 'Victor, you know the lock screen password, don't you? If you don't want to do push-ups why don't you just push in the password?' I have a very, very smart solution for that," Sonck explains. "I basically spam backspace 20 times a second […] to remove anything I might try to fill in until I did my push-ups."
Full project details are available on Sonck's YouTube channel, while source code has been published to GitHub under an unspecified open-source license.