Thread Tools
This thread is privately moderated by Jack Crossfire, who may elect to delete unwanted replies.
Apr 09, 2015, 10:57 PM
Registered User
Jack Crossfire's Avatar
Thread OP
Discussion

More path following notes


Free nuggets about lane following are few & far between. There are many nuggets about line following, but not lane following. This story was fascinating for 2 reasons

http://www.roboticstrends.com/articl...riving_in_1995

It was work my generation did when it was in college, with the tools available during our time. It was so primitive to modern eyes, yet it was the bleeding edge for someone living in that time. There was no GPS. Capturing video on a computer was nearly impossible.

The full text of their path following algorithm costs money, but there's a rough description

http://www.cs.cmu.edu/afs/cs/user/tj...haa/ralph.html

The key to their algorithm was a scanline intensity profile. It was done on a 30x32 greyscale image, on a 486. The scanline intensity profile was derived from a test of lane curvature. The lane curvature can be neglected, since the current issue is a straight path.

They captured the entire lane width, converted the trapezoid shaped path to a rectangle by widening & shifting farther rows horizontally. When making the rectangle shape, they made several images with the farther rows shifted left or right by different amounts. Then they summed each column in the rectangle image. The adjacent columns had the maximum differences when the shifting of rows matched the path's true position.


The sums of each column made up the scanline intensity profile. The scanline intensity profile when the vehicle was centered in the lane could be compared to the current scanline intensity profile to give its lateral offset. The current scanline intensity profile was iteratively shifted left or right until it matched the centered one.


They claimed better results this way than with edge detection. The key advantage was immunity to shadows, relying only on visual features running parallel to the road. This method does require training the algorithm with known scanline intensity profiles for a centered vehicle on different sections of road. There was another issue of cropping the image to where the path should be.
Sign up now
to remove ads between posts


Quick Reply
Message:
Thread Tools