Vision Research
Susan & Kino's Intelligent,
Most Excellent Robot
SKIMER
is used to test out ideas in visual navigation. Basically the goal was to see if
one could build a cheap version of the CMU driver ALVINN that learns to drive
by watching a human driver.
SKIMER
uses a lot of memory to implement a visual connectionist scheme called WISARD.
SKIMER is not so much programmed as trained. The user drives SKIMER through a
task and it associates visual images with commands. When SKIMER sees an image
it classifies the image and executes the command that is remembered.
SKIMER
is a simple model of a visually reactive pilot. Higher levels would select
modules that had been trained for specific tasks. At one DPRG meeting we
trained SKIMER to go around the room using its vision. Then it met Rogers
D-BOT. It was confused for a second, but then decided that it shouldn't run
over a fellow Bot and continued its task, even though it had no previous
experience with other Bot's. It continued its traversal until it reached its
goal.
SKIMER
has a Camcorder as the head and only sensor. The image is captured by a
AT&T Targa frame grabber from a previous project. The video also goes to a
short range UHF TV transmitter. The CPU is a 386/40 DX with 4 Meg of RAM and 60
Meg hard drive specially mounted for shock resistance. The base is a six-wheel
all terrain toy like D-BOT.
Although
SKIMER was not the fastest robot to complete the DPRG test, it was the only one
to use the visual markers present. Within minutes SKIMER can be retrained for
different tasks.
Visual Cortex
The
Visual Cortex system is a project to do color based object and face tracking as
the first stage of an android system. It contains multiple tracking algorithms
that can work in concert to select the focus of attention or when the android
should look. It should work with any Video For Windows devices like USB
cameras.
Viscor1
= Shows how to connect to the camera, and which image to show. It uses Windows’
video system to select the camera/video device. Once selected, open it and the
system starts working. After that you just close it. The color view is just the
raw video input with a grid on top and circles to show where the 'target' is.
The interest view shows a transformed color image that shows skin color. I trained
the system on lots and lots of skin versus background images. The focus view is
just the grid and the ball, with the final segmentation in the upper right.
Viscor2
= Shows how to setup the command grid. The image is the 'Interest' view and
shows skin-like colors vs. non-skin-like colors. Here is where you define when
each command is sent. If you press default it sets up the grid seen above. I
did it this way so if the head is different you can make adjustments.
Viscor3
= The Comm port setup page. I left it simple text input. The image shows the
'Focus' view, with the segmented image in the upper right. Message is the
command that would be sent. It is 'ss' since the target is in the center. I
will filter multiple characters out.
Viscor4
= Dialog page. Image is of an 'Interest' view. It contains a web browser, and
connects to a local Alicebot server running an ANDY script. I can modify the
script as needed and the vision system can update the script dynamically.
Links
CMU Computer Vision Home Page Since 1994 a
central source for links relating to computer vision research.
CAMSHIFT: Computer Vision Face
Tracking A method similar to the tracking of Visual Cortex developed
by Intel.
AMP Face Tracking Project A method similar to
the probability component for VisCor under development at CMU Advanced
Multimedia Processing Lab.