Monday, September 11, 2006

It's Just Like Driving Miss Daisy... With a Robot.

A chinese car company has just recently built a car it claims can effectively drive itself. It currently is limited to 37mph but can get up to 93mph before becoming "confused." Maybe it's me, but the fact that we're using the word 'confused' to describe a car just tickles me pink.

While Chinese companies have a pretty bad history of blatant lies about their technical achievements, I hope this is real. I hope because I'm immensely interested and am DYING to see how they did it.

DARPA recently had it's grand challenge where the Volkswagen Touareg from Stanford University, approriately named Stanley, won the day. Technologies such as radar and laser scanning were used by almost every team, which I found silly. There was one team that attempted binolcular vision, but their Toyota pickup truck had a tendency to flip out and start killing people.

The thing that puzzled me was the pervasive use of laser and radar. I considering binocular vision to be the obvious choice, but not in the way these guys were doing it. This is where a psychologist would come very much in handy, namely, we can't imitate animals until we understand them.

As best as we can tell right now, the primary visual function of differentiating between stimulus objects is based on contour. Our brains seek out lines of contour to figure out where's what. That's why camouflage works best when it's a seemingly random micture of varying colors and dots. It disrupts the flow of contour lines, and blends whatever is trying to hide in with the background.

The only team to try the binocular vision created a difference map on the captures images, analyzing for areas of no difference and where there were sudden changes in color or intensity. They then told the vehicle to follow

For example, open up Photoshop, or any good graphics program, for that matter, and go into filters. You can select 'Highlight Contours' or something like that. It's instantaneous. All a computer has to do is find lines of contour, and whatever is an enclosed line is a single object. That image is than compared to the second camera's image to determine distance. That object is then assigned a temporary id and tracked on the screen.

It's much easier to tell the computer how to interact with objects instead of just arbitrary numbers in a massive difference map or a gigantic array of coordinates from radar and laser. You can supply a shape for a person, dogs, trees, or other cars. You can supply information on what to do if the shape is changing position or shape. You can turn the information coming in into a giant, 3D video game in which the car interacts. It's processor-heavy, but throw a few good computers at it and it's nothing too big.

I would assume there was something drastically wrong with my logic if some, ANY of the teams had mentioned trying this approach. Even in passing. Not one did.

Look Out: Chinese Send Unmanned Vehicle Onto Streets (Via

No comments: