Peeper Pam, she’s a Computer Vision

We consider ourselves to be AI sceptics, so when the nice people at Raspberry Pi sent us their new (70 dollar) AI kit for an early look, we weren’t sure what to make of it. After a look at what it actually was, we thought ‘ooooooh, it’s a Computer Vision thing’.

Computer Vision like the quiet member of the AI family. It beavers away on problems, getting stuff done while its more popular siblings make a lot of noise, doing parlour tricks and shouting ‘Aren’t I clever?’.

Learning by doing is the only way things make it into my brain. Here’s a simple proof-of-concept that shows an example of what you can do with a few cheap components, and the weight of Open Source behind you.

The Idea

The idea is simple. Getting a computer to monitor a video stream, apply a model to the frames and send an alert to a simple single-purpose device when the model thinks that the video contains something you’re interested in.

Depending on the model that is examining the video it can be anything from

• a human,

• a human in a certain pose a human with a specific gait,

• or a human with a specific gait who’s just fallen over and is carrying a banana (you get the idea).

As a first stab at making something, we’ll just make the mere presence of a human trigger the remote device. Never being ones to shy away from turning serious ideas in flippant implementations, it was decided to name this first prototype ‘Peeper Pam’. Pam’s purpose is monitoring an area outside of your field of view for human activity (disclaimer: Pam will not protect you from Alien Invasion).

The Device

Computer vision tends to classify things with probabilities attached (eg I’m 69 percent sure that’s a banana) so the device should provide and alert and a probability. To do that we’ll pop a light on it and an old fashioned analogue needle dial, mainly because we think they look cool and there was one lying around. The important thing is to display information in a way that gets your attention and also conveys a measure of how certain the model is about what it has seen.

The extra components we needed to make this device were inexpensive (~$15) and they’re all described over at our GitHub page, where we’ve also shared all the code we made to get things working.

Video

Now that I’ve told you why there are serious ideas behind all this stupid, here’s a link to the video we made that better describes what we did: