VASC Seminar Announcement
            =========================

Date: Monday, 10/4/99
Time: 3:30-4:30
Place: Smith Hall 2nd Floor Common Area

Speaker:  Vladimir Brajovic
          CMU Robotics Institute
          http://www.cs.cmu.edu/~brajovic


Title: High Resolution Dynamic Triangulation Sensor for Rapid Range Imaging

Abstract:

In this talk I will describe a new computational sensor for triangulation-based 
range imaging.  The chip has been recently designed and is now being fabricated. 
I will describe the functionality of the sensor and show what performance we can 
expect from it in another couple of months.

In most industrial range imaging applications, active light-stripe
triangulation is found to perform best. It gives the best compromise between
accuracy, speed, system complexity, and cost. A sheet of light is
projected onto the scene and is viewed by a camera from an oblique
angle. The intersection of the light sheet and the object creates a
light stripe contour indicative of the object shape at that location.  

In its naive implementation triangulation is slow. First, a light stripe is
positioned at a particular location on the scene. Second, a CCD image is
captured and processed for the location of the light stripe in order to
obtain one slice of range data. Then, the light sheet is repositioned and
the process repeated to obtain another slice of range data. In more
creative implementation multiple light sheets or light-stripe patterns are
projected to obtain multiple slices of range per each CCD frame. 

The fastest light-stripe triangulation method is so called dynamic
triangulation [Sato, et al., 1987]. Instead of discretely repositioning the
light stripe, the method rapidly and continuously sweeps the light sheet across
the scene. Each pixel in an imager needs only to detect the TIME when it
sees the light stripe. From this time, the angular velocity of the light
sheet and the geometric parameters of the setup the range map is computed.
Theoretically, hundreds of frames per second can be captured with this method.

Unfortunately, CCD cameras are too slow to take advantage of the dynamic
triangulation. Custom computational sensor are required and have been built 
by several groups including our lab [Gruss, Kanade, 1990]. All of
these sensors are cell parallel--each pixel is concerned with its own light
signal and has its own time memory. 

The device that I am going to show takes advantage of the fact that there
is only one pixel in each row that is illuminated by the light stripe at
any point in time (this was observed by Kuo and Carley, CMU, in 1992).
This leads us to the row-parallel architecture. One-per-row circuitry
detects the location of the illuminated pixel within each row. The pixel
identity and the timing WHEN it saw the light are paired by recording
address-time information in one-per-row memory. The fact that the time
memory is taken out of the pixel leads us to spatial resolution 10x greater
than any of the previous devices reported. 

The current prototype will produce 64x64 range maps at 30-100 fps. The
circuitry, however, is designed to support the ultimate array size of
256x256. If the current device works well, it will be only matter of
imprinting more pixels in silicon.



Vladimir Brajovic is a Research Scientist at the Robotics Institute 
where he leads the computational sensor research.  He received his PhD 
from CMU in 1996, studying under Prof. Takeo Kanade.