Yesterday I went to the danish exhibition that is featuring Stanley - winner of the DARPA desert challenge some years ago (2006 maybe).
Now that's an awesome robot with some fancy equipment! I think it had 5 or 6 laser range finders on the roof constantly scanning the profile of the road in front of it and mixing this together an image from a camera on the roof of the landscape far ahead in order to recognize flat surface for driving on. It is based on some sophisticated machine learning algorithms so that it can adapt and learn from what it sees with its sensors
Reminds me of this: Recently I saw a "parking-guy" (dunno what they are in english, ticket-dude, the one nobody likes) - He was measuring the distance of the sidewalk/ next intersection and a parked car… With a laser.
As far as I know, laser-scanners for robots cost trillions… But heck - if a parking-dude can get one, then why should my robot settle for ultra-sound?
Sounds good! Are you willing to share your thoughts on the image recognition in a blog entry or will you save it for when you have something ready to show?
I think you’re right about laser scanners being really expensive!
That thing IS pretty damn That thing IS pretty damn impressive. I’ve been wanting to play around with cameras and image recognition. I was thinking a relatively easy, introduction project might be to put a camera outside my house and develop a system to tell me when a package delivery guy is outside. The various delivery trucks (in the US at least) are big and rectangular and each company has their own distinct color scheme, so it seems like it wouldn’t be that hard to look for a big brown, or white/purple, or yellow rectangle in the part of the picture that represents the street.
I think maybe you need a little more than edge detection unless you’re going to only use it with very different colors. I have tried edge detection (various algorithms) before and they worked fine for synthetic images but when it was about natural pictures it was a mess of edges. They only seemed to be something you would use in a pipeline of operations, where you might also have segmentation based on color and other stuff.
The stereo imaging would probably need some correlation of some similar points in the two images although I have never tried it so I’m not too sure about it.
I’m going to build a lego platform where I can put my laptop and a webcam and then hopefully it’s hello to image recognition.
