We have tried pressure sensors, ultrasonic sensors and they do work but it gets complicated when deploying on the field and with customers.
wonder if when deploying with customers we can put a tiny camera on post facing our tanks and some tiny ML to tell liquid 75% 50% and 25%.
what do you think?
Hi,
So this is definitely an interesting use case. Some ideas:
Visual regression is potentially a good option here, basically you select ‘Regression’ as the NN block and train to output a single value corresponding to how full the tank is. We have some work in progress on this topic doing weight estimation from an image but there are some considerations - The best case scenario for good results is an image where a single container is centered in frame. Additionally this would generally only detect the level in a single container. Detecting multiple containers at once in the same image gets a lot trickier.
On the other hand, if you only want to detect multiple containers above or below a fixed threshold, you could run object detection where only, for instance, empty containers get detected and partially full containers are ignored.
There are definitely more advanced workflows you could imagine, including some custom CV processing. but these two are I think the simplest places to start with the existing impulse designs available in edge impulse.
references:
Best,
David
@buton see also Vision to Mass Impulse Calculator and a blog post about weight estimation using visual regression should come out this week on our blog.
Another option, maybe even easier as @daschwar also mentioned is just good old OpenCV. Assess height of the tank, look at how high the pink goes, done.
