I’m working on a lightweight computer vision algorithm for Robotgrrl’s Robot Missions project, and I’ve been exploring neural networks and learning machines on low-power hardware. This is a simple regression tool that breaks an image down into 64 samples and tries to determine which pixels are Jason (blueish) and non-Jason (reddish). It works surprisingly well in ideal situations!
The end goal is to create a feature identification tool that could help a robot navigate around rocks or branches or penguins.
The status feature also seems to be working correctly. (I may have hardcoded that)