Date: Tue, 27 Jun 2006 01:28:30 -0500 (CDT)
Subject: more road segmentation
X-UID: 179

Here is another test of the road segmentation.

http://golem5.org/robot1/video/roadseg20060626.mpg

The road is indicated by solid black.

The blue line indicates the farthest point of the road from the bottom of 
the screen. Note that the hood of the car and a reflection of the 
dashboard causes problems at the bottom edge of the frame. This forces the 
blue line over to the right sometimes.

The green line indicates the center of mass, left to right, of all 
segmented pixels. This tends to remain in the center of the frame as the 
driver has good lane position. Were the driver to drive onto the right 
shoulder, the green line would shift to the left.

The blue line tends to indicate turn direction. The green line indicates 
good lane position. This is probably enough for the robot to follow a road 
now. If the key patch chroma histogram were to suddently change, it would 
also know that it had driven off the road.

An excellent idea a DPRG member suggested is a longer running average of 
the key patch histogram. This way, the robot does not lose track of the 
road segmentation it has locked onto if the key patch suddenly changes. 
This is computationally cheap. In fact, everything done so far is 
inexpensive.

The strange reddish aurora in the sky is due to mapping from YUV space 
back to RGB. All of the image segmentation is done in YUV space. 
Generating the pixmaps requires conversion of YUV (which is really 
spherical coordinates) back to RGB (cartesian). Either I have fudged 
something or this is numerically ill-conditioned. I'll have to look at 
that.

The blockiness of the road segmentation is due to using the same code that 
will go on the robot. In order to maintain a reasonable frame rate, the 
network camera on the robot is set to lowest resolution which is 160x120. 
So all road segmentation computation is done at this reduced resolution. 
When making these movies, this low resolution image is then upsized back 
to 320x240 to match the originals from the digital camera. I could make 
movies that look better. But they wouldn't represent what the robot will 
be doing.