Date: Sat, 24 Jun 2006 23:09:50 -0500 (CDT)
Subject: road segmentation with YUV chroma
X-UID: 178

Road segmentation using simple two chroma channel histogram thresholding 
(YUV space, not RGB) in a circular image buffer works well enough. It's 
good enough for the robot. It just has to be integrated with the rest of 
the control system.

http://golem5.org/robot1/video/roadseg20060624.mpg

The original images came from a digital camera. It was on an upside down 
Bogen tripod that I held while walking down the path. The weight of the 
tripod helped to stabilize the camera and prevent it from bouncing too 
much with my walk.

The upper image is the original while the lower is the output from 
CircularImageBuffer::outputRoad(). There is very little source code for 
this. Not counting the third party open source libraries like jpeglib, the 
entire onboard robot control and image processing system will probably fit 
in under 2000 lines of code. Only a few hundred lines of that will be for 
vision.

All black means road. The blue line indicates the end of the road. It 
approximates the direction and speed the robot should take. This will be 
the output of the monocular vision system.

The road segmentation key area is a small rectangular patch at the bottom 
center of the frame. This is presumed to be the road. The rest of the 
image is compared against this key in YUV space, ignoring the Y intensity 
channel. This works well enough although shadows are still a problem. Note 
that the road segmentation is never really accurate when there are any 
shadows.

Note that no use of the Y intensity channel is used. Everything is done 
entirely with the U and V chroma channels. To address shadows properly, I 
suspect that information in the Y channel is necessary.

I've also tried running old data sets from previous robot drives through 
the system. They are far more problematic. One set was done just after 
dawn and is underexposed. The image lacks sufficient dynamic range for 
histogram thresholding techniques to work well. The other is done in the 
middle of a parking lot with drivable surface boundaries formed from 
concrete instead of green grass. The common factor is insufficient dynamic 
range in the image chroma histogram for easy segmentation. That can be a 
result of available light and the camera sensor or an environment in which 
there is insufficient difference in color.

I believe that good field vision systems are possible. But they are not 
easy. It is like chess computers. The concept is simple. Implementation of 
grandmaster level systems requires significant resources with optimized 
search. A vision system is similar in that it must use many techniques and 
somehow evaluate a best guess.