Office: NSH 3124 or the Intel Lablet at 417 South Craig Street
Phone: (412) 605-1401
I recently began work at the CMU Robotics Institute and the nearby
Intel Research Pittsburgh Lablet on extending the capabilities of
sensor networks through the use of autonomous robotics. At CMU I
work closely with Illah
Nourbakhsh, among others. At Intel I work closely with Rahul Sukthankar, among others.
My research interests include autonomous robotics, distributed
systems, sensor networks, and visual navigation.
As a warm-up exercise in advance of more substantial research, and as
a demo for the Intel Lablet's November 2003 open house, Rahul
Sukthankar, Illah Nourbakhsh, and I implemented visual odometry
and obstacle avoidance capabilities on a camera-equipped Palm Pilot
Robot Kit. We demonstrated:
precision distance measurement to better than 1% accuracy under
conditions of extreme wheel slip
visual servoing to maintain a straight course in spite of differential
drag on the robot (accurate to within 3-6% of straight over several
feet of travel)
reliable detection of any vertical drop greater than 1.5 inches, and
the ability to stop the robot safely in advance of any such precipice.
All of these were implemented via optical flow information derived
from a single robot-mounted webcam. No encoders were available on the
robot, and no other sensors were used. A
poster and two videos demonstrating precipice detection
and distance measurement are available.
Automatic Stitching of Multiple, Oblique Camera Views
(November 2003)
Also as a demo for the Intel Lablet's November 2003 open house,
Yan Ke,
Rahul Sukthankar, and I implemented a demo of new
keypoint-matching improvements developed by Yan and Rahul. In this
demo we show four uncalibrated cameras viewing a common scene. Common
keypoints are extracted from any camera overlap and used to define
corresponding intercamera homographies, and then those homographies
are used to stitch together a single mosaic containing all the camera
views. The demo runs is near real-time, with an update latency of
approximately 10-15 seconds.
The poster and photos of the demo will be available shortly.