i’ve already successfully used tracking with vvvv, but this one is still too hard for me…imagine a long corridor where people are walking around, ok to track, but what to do if i only want the information (size,x,y) of the upper part of their body, even better the chest?
once i used the color tracker after using the background supression- the c tracker gave me a lot of points for every object, so i sinply took the middle of all…obviously this doesn’t work this time as i need a more accurate result…
that should somewhat be possible with the DetectObject node. maybe opencv even ships with training data for what you need. else you always have the possibility to create your own traindata. has anybody ever tried this?
i assume you want to track both male and female clothed chests.
for the reason of manifold possibilities of clothing i furtherhin assume that this is almost not possible in a direct way.
but as a workaround you could track heads with DetectObject (with the known distributed traing file) and just “transform” your selection a bit “lower”.
Yeah, it sounds to me like your best bet would be to grab the head location with detectobject and then just estimate the torso area based on the head location.
I’ve tried to find information about building your own training files, but haven’t been able to come up with much… if anyone has any links I’d love to see them.
hmm; i’ve had a look at those training files already, but didn’t get a clue either…any help or examples really welcome…
i thought detect objects was looking for faces, not heads- so it works from the front and the back? that would be quite cool already. though, if i could really detect the chest i could also get information if the tracked person is turning 90 degrees, which would be quite cool for my aim; more general, to get the width of the person at the height of the chest.
this is also the first idea i had: to use something like contour and search for the maximum horizontal distance of contour points, which would be the shoulders (most times :); problems here are people viewed from the side, so i thought about a mix of max hor distance for x and something about 3/4 of the max height for y…what do you think? still stick to detect object?
writing about it i realize that my solution might work for people already in the wiewing area, but lacks while new people are coming in…how does detect object solve this? what if the head is the first to see for the camera, would it detect some round object without head until the body can be seen? sorry if this is a stupid question, didn’t use it yet, no cam/vids here…
detectobject with the face-trained file will not work for heads from the back. it is actually looking for things that look like faces.
opencv ships with the following info in \data:
the folder contains trained classifiers for detecting objects of a particular type, e.g. faces (frontal, profile), pedestrians etc. Some of the classifiers have a special license - please, look into the files for details.
there are actually files named _lowerbody, _upperbody and _fullbody. that maight already be what you are looking for. get your copy of opencv here.
\apps\Haartraining includes an application + description how to create your own training files.
