|
As a graduate student at Ohio State in the mid-1970s, I inherited a unique computer
vision laboratory from the doctoral research of previous students. They had
designed and built an early frame-grabber to deliver digitized color video from a
(very large) electronic video camera on a tripod to a mini-computer (sic) with a
(huge!) disk drive—about the size of four washing machines. They had also designed
a binary image array processor and programming language, complete with
a user’s guide, to facilitate designing software for this one-of-a-kind processor. The
overall system enabled programmable real-time image processing at video rate for
many operations.
I had the whole lab to myself. I designed software that detected an object in the
field of view, tracked its movements in real time, and displayed a running description
of the events in English. For example: “An object has appeared in the upper right
corner . . . It ismoving down and to the left . . . Now the object is getting closer. . . The
object moved out of sight to the left”—about like that. The algorithms were simple,
relying on a sufficient image intensity difference to separate the object from the
background (a plain wall). From computer vision papers I had read, I knew that
vision in general imaging conditions is much more sophisticated. But it worked, it
was great fun, and I was hooked.
A lot has changed since! Dissertation after dissertation, the computer vision research
community has contributed many new techniques to expand the scope and
reliability of real-time computer vision systems. Cameras changed from analog to
digital and became incredibly small. At the same time, computers shrank from minicomputers
to workstations to personal computers to microprocessors to digital signal
processors to programmable digitalmedia systems on a chip. Disk drives became
very small and are starting to give way to multi-gigabyte flash memories.
Many computer vision systems are so small and embedded in other systems that
we don’t even call them “computers” anymore. We call them automotive vision
sensors, such as lane departure and blind spot warning sensors. We call them smart
cameras and digital video recorders for video surveillance. We call them mobile
phones (which happen to have embedded cameras and 5+ million lines of wideranging
software), and so on. |