With the simple addition of about $2 worth of parts, an ordinary video
camera can be transformed into a range-based segmenter. This has a
number of uses in human-computer interfaces as described below. The
device works by flashing an invisible infrared LED on and off, and
then looking for which pixels got brighter. The inverse square
fall-off of diffuse reflection ensures that only objects relatively
close to the camera will change in intensity. While there are other
ways to segment objects from the background (e.g. stereo, motion,
color) none of them has such a simple hardware configuration and such
minimal processing (just frame subtraction).
The demo system we have built uses an off-the-shelf Radio Shack surveillance camera with built-in IR emitters. For the system to work it is important that the camera NOT have an IR-cut filter. However, such a filter is an extra expense anyhow and is usually not included on commercial camcorders (try taping your TV remote control). To complete the system we modify the basic camera to put the IR LEDs under computer control, and add a visible-cutoff filter to increase the signal to noise ratio (somewhat surprisingly, the basic video looks the same). IBM's CVM frame grabber & DSP board is used to process 256 by 240 pixel images at 7.5 Hz without intervention by the host processor. This includes a local averaging step (4 arithmetic operations per pixel) to clean up the raw difference image. We have also ported the system to use the main CPU of a workstation or laptop computer at full frame rate. This version uses a modified camera which automatically flashes a number of infrared LEDs and returns a standard NTSC signal which is digitized by a USB adapter. The camera is especially suited to automotive applications because:
|
| Contact: Jon Connell | Last updated: 6/12/02 | ||
 |
 |
 |
 |
