Home
The aim of this software is to be able to count the number of people present in a crowd, especially in protests, and provide an estimate as accurate as possible. This software is currently at a very early stage.
Note : The result video does not reflect the speed of the algorithm at all ! The algorithm is a lot slower right now but there is a lot of room for improvement there.
You need OpenCV to compile this software (that means the packages libcv-dev and libhighgui-dev). I am trying to keep it in sync with the latest ubuntu stable distribution. It should work well with debian as well. It supposes the include files are in /usr/include/opencv and the libs in /usr/lib. The Makefile is very straightforward to change.
Once done, simply typing :
make
should compile both cvplayer and headcounter.
cvplayer is a simple test utility to check that your opencv configuration manage to read a specific file
headcounter is the main program. headcounter -h gives you instruction on how to use it.
Here is a complete movie of a protest that happened in Lyon (France) on 2010-10-04. Only 720p but hopefully the next one will have better quality :
I also used a cropped version of the "crowd_run" video available there : http://media.xiph.org/video/derf/ Cropped version : http://quemener.yves.free.fr/projects/headcounter/Marathon_crop.avi (267 MB)
I have been unable to find another good example file of a crowd where faces are visible and not too small. Please do not hesitate to contact me if you have such files you are willing to share !
I am using openCV's face detector (haar cascade classifier with the cascade file included in opencv) to detect faces. From there I associate close faces on successive frames in order to uniquely count individuals. There are several drawbacks on this implementation :
-
it is slow
-
it needs a face that is at least 20 pixels wide
-
it works only for people really facing the camera (no profiles)
-
it works bad with caps, hair covering a big part of the face, glasses...
-
it can count multiple time a single person that gets temporarily hidden behind someone else.
-
Possibility of creating a cascade file that would work for smaller faces ? It sounds like it would be feasible. The file given with openCV is optimized for the tracking of a single person face, on a webcam. The 20 pixels size is probably more a choice than a hard limit.
-
Possibility of optimizing the haar detector to make it closer to real-time performances ? I have seen some detectors running a lot faster than this one. There may be room for improvement.
-
A completely different algorithm could look at interest points and optical flow. It would cluster the points moving in a coordinated fashion and try to differentiate individuals this way. This can probably work exaggeratedly well for the marathon video, as the runners shirt have a lot of interest points. It would be interesting to test on a more compact crowd.
You can contact me at quemener dot yves at free dot fr.