During this fun run I have built a set of tools for image recognition with a trained neural network. Hopefully we can integrate the solution to the live environment and test the AI against real life 🙂
The steps I took:
1.Dataset Collection:
I have built a setup using a raspberry PI with a webcam, running a Python script which is making an image every time when movement is not detected.
This approach made me possible to collect a huge amount of images from different locations, which are ideal for training a neural network.
After collecting 20K+ images I have used all kind of methods to sort them and place them in a folder structure which allowed my setup to parse and teach them to the neural network.
2.Training the model:
The model training took me about 4-5 hours, with multiple adjustments and fine tuning. It was getting better after each iteration and finally I have reached the 99.8% Confidence level. 🙂
I have to add that I used my monster PC with 128GB RAM for the task.
3. Making it ready for deployment:
There is a not really well known branch of devices called NVIDIA Jetson, these little guys have an GPIO’s and a GPU cabable of running TERAFLOPS of AI operations per second utilizing CUDA cores!
So I got one of these, and after couple of hours trying to bridge the canyon of differences between my Windows setup and the Linux running on this guy, I have managed to make him the perfect host for my model.
🙂
Now he can tell the PLC if the tote has got a Lid, products, or it’s empty with 99% confidence, without expensive and ridiculous lasers, sensors or mechanical clip-clops 😀
Final Setup 
