count the number of standing people in the view of an overhead camera

Hello, I am planning to do a real time head detection with hough transform algorithm and then count the number of detected people in the scene. In this case the camera/webcam is mounted to the ceiling of a room and people (max 8 people) are standing under the camera. I would like to get any suggestion regarding this topic.

appreciate it. thanks

Answer 1

[Image Analyst]

See papers on crowd estimation here: http://www.visionbib.com/bibliography/motion-f745c1.html#Human%20Activities,%20Crowds,%20Lots%20of%20People

A simple, but perhaps not super accurate, way would be to get an area fraction of the "empty" scene that is covered by people and turn that into a count knowing the area that a single person takes up.

[amir]

Thanks for reply. Actually I have referred to many journals before regarding the crowd estimation but they mostly detect the moving objects. I know this is not an accurate algorithm but there is any method that I can improve it?

[Walter Roberson]

See http://www.mathworks.com/matlabcentral/answers/39545-detect-a-human-figure-in-a-live-video#answer_49322 and expand the comments.

[Image Analyst]

Well pick one that doesn't rely on moving people, if that's your case. I know they've used it for still crowds, like to estimate the number of people at events and demonstrations, like in Washington DC. It became a heated argument because one group wanted to have a million people and estimates showed only 400,000 which made them angry because they needed a million to match the name of their event. Since that event, the National Park Service will not give crowd estimates anymore.

http://www.nbcnews.com/news/nbcblk/black-men-gathering-million-man-march-20th-anniversary-n442226

"....The National Park Service estimated the attendance at the original 1996 march to be around 400,000, but subsequent counts by private organizations put the number at 800,000 or higher. The National Park Service has refused to give crowd estimates on Mall activities since...."

Anyway, why can't you do a calibrated area fraction like I suggested? Post your image if you want.