Just coming back from France but I’m already looking for the upcoming Eurobot contest which will be held in Germany next year… My contribution to this year’s edition was to develop some kind of software that could be used on the robot to detect and localize play elements as well as bins. Even though we failed to get the robot homologated, we learned quite a lot.
Such a thing I would like to apply would be to speed up the thresholds-finding procedure that need to be run almost each time the robot is moved to another location. This was caused by the lack of a proper thresholding algorithm that would adapt to minor illumination changes.
So there is the purpose of this post: I’m now planing next version of my software and would like to include such a feature to speedup last minute settings. Here’s some relevant topics I picked up in a book from E.R Davies:
- Variance-based Threholding
- Entropy-based Thresholding
- Maximum Likelihood Threholding
- Local Thresholding Methods (splitting the image)
Now how applying this stuff to my particular problem… these subjects tend to be more about how choosing a threshold rather than adjusting one dynamically over time. Furthermore, it’s more about finding a threshold for grayscale images that best fits all of its regions than actually isolating each of the color we are looking for.
I know this is a common problem for all teams but I think something simple based on a statistical approach could be tried here.