CraigMattson.net

Solving Cat Problems and Watering Systems with a Raspberry Pi and Machine Learning - Part 1

13 August 2021

I have a son who is now 11 months old and other family members that are allergic to cats so when a neighbourhood cat decides to defecate on the lawn, not only does it smell awful but it poses a number of health implications too. Some of the cats have also taken to climbing up the fence and jump around the roof - waking up everyone at 2:00 AM in the morning. Because it's more than one cat, it's hard to know whose cat belongs to who and the local council wants some $70.00 per cat trap to hire - and that's when they're not all out on loan anyway.

This got me thinking about the ever-growing pile of Raspberry Pi's that I had in my draw for several years. I hadn't really got into GPIO's beyond a couple of 'light an LED' scenarios - I never really had any projects go 'in-flight'. There's always the idea of building a set of Flight Controllers but with various life changes that project has been on the backburner for some time. I also hadn't had much of a look into Machine Learning beyond a few image recognition proof of concepts as I had no real use case for it, until this problem.

One thing cats really hate is water. I figured that if I could get a remotely controlled Solenoid, activated by the front door camera if I can detect the presence of a cat, then this might be a more humane way to scare the cats off. It'd of course only need to trigger for a few seconds until the cat has gone - and using a Raspberry Pi might be well overkill for it, but offloading the capability might reveal a good result.

The shopping list

Back in April, I took stock of the parts I had. Several Raspberry Pi Zero's (that were to be used for the controller project), some Raspberry Pi 2's and 3's,  a Raspberry Pi camera, some crimpers and that's really about it. I jumped on Amazon and Core Electronics website to buy some parts. By the way, if you're following this to build your own - I'd strongly recommend not just buying everything in the list - I really didn't plan this part well and probably bought some unnecessary equipment here.

This would be a bit of a stock pile such that I could begin working on "something".

The first build, kind of...

The package of miscellaneous parts arrived and I got to work stripping some wires, putting some crimp terminals and spades on the end of wires from some surplus auto equipment I had lying around. I've also connected some cameras up to get started with some kind of ML. My desk looked like the below for a few days.

Figure 1 - Some Pi's with Camera Cables
Figure 2 - Wiring up the Relay and Solenoid
Figure 3 - GPIO Pins firing with some basic Python code.

I'm mildly excited at this point - having had to fight Python and dependency issues, I've managed to get a small script to turn on and off some pins working on the dreadfully slow Raspberry Pi Zero. I mean, who knew that having Python 2.7 and 3.5 on the same distribution would cause as many issues (do I use pip or pip3, python or python3?). There's a certain sense of achievement you get when you manage to control a physical object. Perhaps it's the sound of a relay clicking and a solenoid *thump* that makes all the difference. 

from gpiozero import LED
from time import sleep

hose = LED(17)

while True:
    hose.on()
    sleep(1000)
    hose.off()
    sleep(1000)

The above script won't win any prizes for creativeness by any means, but that's really how simple it is to start turning stuff on and off. I'm not exactly a fan of Python and I dislike YAML for similar reasons. When you get used to C-style syntax, the mark up makes it pretty easy to place code the way you want to see it. It's fair to say I've developed my own preferred styles for laying out code so moving to a different language that also removes some of these "features" can feel unpleasant. But this is where I think I made a mistake.

I figured that I've been working with Typescript for around five years now, surely these examples are available in Node.js. Of course, there absolutely were examples - there just weren't very many of them compared to Python. When you start going down the Machine Learning path on the other hand, things get a lot murkier and in some cases you end up with Javascript wrappers for Python scripts of which are wrappers for C libraries themselves. This doesn't strike me as robust enough to have such a chain of possible failure points. To add to it all, the Raspberry Pi Zero is also only Armv6 capable - Node.js stopped supporting this architecture sometime in 2019. But I know how to write a basic express app in Node.js so I gave it a good go.

Before I knew it, I had imported onofflodashfslowdbfs, cors, nodemon and started scaffolding some API endpoints. After all, any logic I have that detects images from a camera will need to pass detections to an API method - this would allow some abstraction to take place. With Docker installed on the Raspberry Pi Zero, I'd eventually build an image based on the balenalib/raspberry-pi-node image.

NB: Which by the way, if you're trying to build containers to run NodeJS on Raspberry Pi Zero hardware, the following snippet will be super useful. There was a lot of trial and error to get this exact package and image for the job...

FROM balenalib/raspberry-pi-node AS base
RUN install_packages make gcc build-essential python3 influxdb

A word of warning, it's not a great idea to build these images on the Raspberry Pi Zero directly unless your idea of fun is to spend 30 minutes waiting for all the dependencies to compile and install. If you don't get your multi-stage builds right, this becomes a lot more fun when you have an error further down in your Dockerfile. As I learned through building these images, they are multi-platform by default so building them on a far superior computer and pushing them to a shared private repository was a really good way to handle the compile time slowdown issue.

Suffice to say though, I managed to build something that would respond to commands like:-

  • /gpio/{pin}/{state}/{timer?} - where pin = GPIO Pin number, state = on or off and timer = how long (0 = indefinitely).

Pretty simple stuff. Or at least I thought it was. Things kind of worked - you need various levels of privilege escalation to get this thing working but perhaps the biggest issue was that I'd just spent a lot of time coding something that really didn't need to be this complicated. But it did work, so I put the image into a Docker Compose script, set to always run and loaded it in a box.

Figure 4 - All of the parts ready to open a solenoid.
Figure 5 - Basic webpage that would monitor my four outputs.

Machine Learning

Before, I mentioned Python having a good set of resources available especially when it comes to tinkering about on a Raspberry Pi. Having powered up the Raspberry Pi 2 I had lying around, I connected the camera to it and ran a few example Python scripts on it to see what kind of framerate you could expect - having not done much with Object Detection beyond a proof of concept project at work with Amazon Rekognition.

The first example I downloaded revealed some success - it wasn't overly fast, maybe 3-4 FPS but it could pick up some things like monitors, chairs and teddy bears. I wish I had kept a photo of this one.

All well and good if I want to use this particular library and run the application on the Pi itself, but it definitely wasn't great at detecting cats nor was it very fast at doing so. If I'm going to start writing some serious logic - it's time to do some learning. Microsoft had recently announced ML.Net having a feature update and I figured instead of learning more Python, I could put some C# skills to use. I stumbled across some examples on getting TinyYOLO v3 running on ML.Net using Onnx. There's a lot of material to cover here, perhaps for another blog post. But if you're starting out in Machine Learning and think reading this one example is going to solve all of your Machine Learning queries, you'd be absolutely wrong.

When I started implementing pipelines, using Neutron to work out how to invoke the model and learning about how to reduce image quality to speed up image detection - it becomes overwhelming very quick, and most of your object detection labels and the way scores are calculated changes dramatically. Different settings start to throw out bounding boxes and even what's actually being detected. None-the-less, I'm only interested in "Cats" and "not Cats" - so I figured this should be relatively straightforward. With my API in place and several algorithms, it's time to test.

Figure 6 - Swagger Endpoint for my new AI / ML API

With so much excitement, I begin uploading some images. By about the fourth image, I was immediately excited that the bounding box is perfectly around the dog, and it detected a dog. That percentage could have been a little bit higher, but who cares - it detected the right thing!

Figure 7 - First detection is a good match for a Dog.

Some more pictures in, and we finally get to a pretty obvious cat from some Aiptek camera back in the day (remember those guys?). The percentage wasn't exactly super high, but you can only see two legs. Maybe that's why ;)

Figure 8 - Obvious cat is a cat.

Alright - time to give it something obvious - two dogs and a cat. What do we get? - bzzt. Nothing.

Figure 9 - Twang - there goes the model. :(

Suffice to say, I went through a lot of different kinds of models to try and determine the best one for the job. Some were slow, some were fast, nearly all of them were only producing < 1 frame per second on non-GPU accelerated hardware. I even had a go at general image classification - and despite only having two kinds of things I was looking for, most algorithms there were basically treating things as "Cat" or "Not Cat" - rather than a whole list of possible outcomes. But - I did have some success so I would proceed to bundling this up in an image with all available models ready for testing.

The RTSP debacle...

It was clear early on that the Raspberry Pi's camera was pretty ordinary. I had however purchased a Ubiquiti Camera - the Bullet G4 and mounted it to the front of house. There's a few cool things that this particular camera supports - RTSP is probably the most important as well as some Motion Detection that I can hook into. All I needed was an RTSP library to get images from the stream (short of writing my own RTSP interpreter - surely I'm not the only person in the world consuming this feed), and I'd pass it onto the detection API before deciding whether I wanted to turn the relay on (and for how long).

Well - this turned out to be an even bigger hurdle than the Machine Learning itself. After traversing through Github and Nuget for libraries, it was pretty clear that RTSP feeds can be vastly different, and most were designed for particular cameras or use cases. In particular, the Eufy camera I had before kept breaking most C# libraries I was trying to use, except for Emgu.CV.

Turns out, this Machine Learning library does have a fully baked RTSP reader and it works fine if you're running everything on Windows. I was not - my servers are mostly Linux and anything Windows are my desktop PCs which are used for a lot of other things (and despite having Solar Panels on the roof for daytime, I don't want to be chewing up that much electricity for the one or two times a night a cat may be detected). Despite following build instructions for Ubuntu, I could never get the library working without throwing many dependency errors.

The Emgu.CV library requires a few dependencies:-

  • Emgu.CV
  • Emgu.CV.runtime.windows
  • Emgu.CV.UI

With these installed, connecting to the RTSP stream is fairly trivial. You create a VideoCapture object, subscribe to the ImageGrabbed event and start the feed. You can hook into various error methods to attempt a retry in case of disconnect but for the sake of brevity in the code below, I've kept it super simple as it's useful if you ever need this, at least in a Windows environment.

var capture = new VideoCapture("rtsp://...");
capture.ImageGrabbed += ImageGrabbedHandler;
capture.Start();

void ImageGrabbedHandler(object sender, System.EventArgs e)
{
    var frame = new Mat();
    var captureRetrieved = Capture.Retrieve(frame);
    if (captureRetrieved)
    {
        var frameImage = frame.ToImage<Bgr, byte>();
        var bitmapImage = frameImage.ToBitmap();
        
        // throw this bitmap whereever it needs to go
    }
}

Unfortunately, this is where the Cat Launcher venture ends. Having installed the camera and attempted some image detections, it's clear the algorithms aren't quite working. In some cases, it was detecting the local neighbourhood kids as 'Cats' from time to time, and 'Cats' as 'Dogs' and various other anomalies. That's not to say all of this is wasted, I'll definitely be picking it up again. But as I headed into Autumn and Winter, the cats don't tend to spend too much time digging up the lawn. Or perhaps the Jack Russell that was staying with us may have helped keep some of the cats away. In any case, until these images can correctly identify a Cat, this project is on hold.

Figure 10 - Cat not found at night.

In the meantime, and for my next post - a new project cropped up in May to automate a gardening system. Having installed a bunch of Planter Boxes to grow some veggies, the same kind of system at least without the Machine Learning component is required. So I've pivoted to building an automated gardening system in the meantime. Stay tuned for Part 2!


Building a new 3D Printer Workbench

27 July 2021

Several years ago, I purchased a 3D Printer due to a growing number of projects I wanted to start building. Given the prices had dropped considerably, I opted to purchase a Da Vinci Pro 3D Printer - a fully enclosed one that is somewhat suitable for ABS and PLA printing such that I could design and print a few cases and models for some electronics projects I was planning. Suffice to say, I got as far as 3D Printing a goose that lived on top of a Christmas Tree in 2019 before the preparations of having a baby began (read: lots of painting, house maintenance and other things during lockdown that were an incredible time suck).

Figure 1 - Da Vinci Pro 3D Printer
Figure 2 - Printed and Painted Goose
Figure 3 - Goose at the top of the 2019 Christmas Tree

By the way, if you don't get the reference - I strongly recommend you check out a game coded and published by a local developer here in Melbourne - Untitled Goose Game. While short, it's a barrel of laughs.

With the garage walls painted, floor epoxied and my shelving in place, it was time to set up my workspace for 3D Printing and other projects. I had a table space of approximately 120cm x 60xm of which the printer itself took up around half of that. Somehow I managed to cram a soldering station and storage on the table in the remaining 60cm x 60cm. I've recently purchased a Resin printer and UV curing station in the hopes of printing miniature models with detail and of course there's no room to place these items in the current setup. This got me thinking about building a new workbench.

Earlier, I had some success building a couple of computer tables with an Alex shelf from Ikea and a Kitchen Bench Top from Bunnings. The end result is aesthetically pleasing and quite practical without costing the world to build so I figured something in a smaller scale might fit the bill.

Figure 4 - Desk setup from earlier 2020 feat. Alex Book Case and Bench Top.

If I were to replicate the same kind of thing as above with a couple of Alex Drawers, some thinner bench tops and some thick heavy duty castor wheels - it would be a super useful bench to move my growing 3D Printer collection and associated spools as well as store my ever-growing collection of electronics and microcontrollers. There's not much wrong with the dimensions of the existing table - 120cm x 60cm should be sufficient for these two printers and the curing tank. This will leave a little bit of space to work on as well, with plenty of storage for the growing electronics collection. And of course, the remaining 120cm x 60cm table can be used for the full soldering station with plenty of elbow room.

The order goes something like this:

The total cost comes to around $380 if you have some screws lying around otherwise add a few extra dollars for some wood screws.

Having purchased much of this before lockdown no. 5 and a spare weekend to get it all done in, I set out to build the two Alex drawer shells and had a production line going for the drawers themselves. No photos of this stage, but having built them before the best thing I can suggest, like any Ikea product, is sort the screws into piles before beginning and have easy access to each of the parts.

The next step was to build a platform for the drawers to live on. At this stage, I found some wood screws and was about ready to start drilling away until I realised I had no washers to hold the castor wheels in place! Not to worry - I've got a 3D Printer. So I pulled out my trusty copy of SketchUp 2017 and drew a countersunk washer. Having put 24 of these on the 3D Printing platform, it's now time to get them printed while assembling the remaining drawers.

Figure 5 - Washers in XYZPrinting
Figure 6 - First castor wheel screwed to the first panel. Only 5 more to go!
Figure 7 - All six castor wheels attached to the bottom platform.

With all six castors installed, the next step is to mount the Alex Drawer shells to this panel. When assembling the shell, make sure you've got the right panel with the holes in it facing downwards as the top panel in this instance is solid. They are interchangeable - just be mindful about placement.

Figure 8 - Alex Drawer Shell on each side of the bottom platform.

Now these drawers aren't perfect fits - there's about a 2cm overhang but for my use, this isn't a problem as the power board will mount at the back some time in the future.

Figure 9 - Panel overhang by 2cm.

Once complete, the next step is to cut VELCRO strips to length in four places to mount the top panel to the top of the drawers. I like the Velcro strips as it stops table movement on the top, but it's also removable if you need to adjust it or to disassemble it later. It also helps offset any minor table warping so it sits somewhat flush on top. I find it easiest to bind both sides of the VELCRO strip together before cutting into lengths. Remove the sticker from one side and place it on the wood panel but leave the sticker on the side that will attach to the top of the drawers until ready to assemble.

Figure 10 - VELCRO Strips in place.

The final step is to remove the protective sticker and carefully flip over before mounting. By slowly letting the front down, I was able to make minor adjustments until both sides were square with each chassis. Once in place, push down where the strips are attached to ensure a good seal with the cupboards underneath.

Figure 11 - Mounting the top panel to the top of the drawers.

Having then assembled the smaller drawers, it's time to insert them into the chassis and screw them in.

Figure 12 - Assembled Drawers awaiting mounting
Figure 13 - Completed drawer installation.

It's finally done - end to end it took around two hours with most of that time in assembling the drawers themselves... and probably another two hours moving all of my electronics and microcontrollers to the drawers. With a bit of cleaning and unboxing the Resin printers, I can see how they'll look including turning the printer on for a quick run.

Figure 14 - Future 3D Printer Home

For $380, this looks and functions fantastic. It really is just a platform with a few drawers attached, but now it's home to most of my project work ready to wheel around the garage as necessary. Given how well this came out, there might be another one made in the coming months to move the soldering station among other things to.


« < 1 2 3 > »