What it is
A Raspberry Pi 5 with a camera module watches for hand gestures and controls my PC: launching apps, toggling Discord mute, adjusting volume, or shutting the machine down. When the PC is asleep, the Pi wakes it with a Wake-on-LAN packet over a direct 3-foot Ethernet link between the two machines.
Pi Camera -> MediaPipe landmarks -> gesture classifier -> WebSocket -> PC agent -> action
|
v (if the PC is off)
Wake-on-LAN via EthernetHow it works
MediaPipe extracts 21 hand landmarks per frame on the Pi. Rather than using canned gesture recognition, the classifier is trained on my own recorded samples, so the gesture set is exactly what I want it to be: fist, open palm, thumbs up, peace sign, and so on. Recognized gestures are streamed over WebSocket to a lightweight agent on the PC that maps each one to an action.
The details that matter
- Destructive actions need confirmation. A fist means shutdown, but only if you repeat the gesture; an open palm cancels anything pending. Accidentally shutting down your PC because you clenched your hand once is exactly the kind of failure mode this design guards against.
- The Pi outlives the PC. Because the controller is a separate always-on device with its own direct Ethernet link, it can bring the PC back from a powered-off state - something a software-only solution can never do.
- Cheap, honest latency. Everything is local: no cloud inference, no round trips. Landmark extraction and classification run on the Pi; the PC agent just executes.
Why it’s here
This is the smallest project on this page and also the most fun to demo. It covers a different axis than the others: computer vision, model training, embedded constraints, and designing a physical interface that has to fail safely.