Description
Uses CamShiftTracking. I always ran it on the post-background-subtracted video. That worked well, but it tended to lose the person when they moved too quickly. I track only using R. I recently added multidimensional histograms so it can use more than one channel but it doesn't work so well. I use motion templates for my tracking now.
On the other hand, I think it should be possible to get this working... cam shift is *supposed* to be good.
Parameters
HistogramDimensions
The number of histogram dimensions; the number of channels to use in tracking.
VMin
Threshold for "almost black."
SMin
Threshold for "almost grey."
These parameters screen out noise. Read "Seeing with OpenCV" from Servo magazine for details. Swistrack uses the library described in that article. (Google for camshift-Servo-OpenCV-part3.pdf.)
InitialWindowSize
The length in pixels of the side of a square to make an initial window size. It will always use the same size. Really it should set this automatically from the particle...
FrameKillThreshold
The number of frames to keep the track around before killing it, if it hasn't been updated.
MinNewTrackDistance
The minimum distance a new track must be from existing tracks.
TrackDistanceKillThreshold
If two tracks get less than this distance to each other, the shorter one will be destroyed.
MaximumNumberOfTracks
The maximum number of simultaneous trackers.