One nine second play is over
forty-five megabytes of input data. Each frame of that sequence must have lines detected
and field features tracked so that a new rectified sequence can be generated. The
rectified frames are 700 by 679 pixels, so that sequence alone is another 128 megabytes of
data. The difference blob detection requires more processing on the rectified frames,
including morphological operations and filtering in space and time. At each tracking step,
a large number of correlations must be performed. Finally, the algorithm usually requires
running multiple trackers at once so that each tracker must know about all the players
nearby. However, the current system is parallelizable and could be designed to run in
near-real-time with some specialized hardware and hundreds of megabytes of fast memory.
Real-time implementation of the current tracker will require systems capable of fast
computation, specialized real-time image-processing hardware, and hundreds of megabytes of
fast memory.
A simplified version of the football tracker
has been used to track people in real-time as they move about an interactive environment
called the KidsRoom.
This system is described in Vismod TR#403. |