Best way to detect full body poses?

Hi there,

I´m in the need of detecting the user´s pose in a scene. I need to detect simple full body poses in which the user is always facing the camera, such as arms opened and legs closed as forming a T, or arms and legs wide open as forming an X. There might be quite a few variants… like one arm lifted straight up, etc… so the solution should contemplate a pose recording procedure.

I have a kinect sensor at disposal.

I did some skeleton tracking with Kinect some time ago, and found it too noisy, so I started to look into OpenCV and the HaarFinder class trying to match the user´s silhouette to some predefined silhouette (everything with black background). But I got stuck generating the haar classifiers (opencv_haartraining.exe ends up consuming my entire physical memory… this might be that I´m either doing something wrong with the samples -positive and negatives-, or that my solhouettes are not suitable for haar recognition).

I´m working with solhouettes pretty much a this one: https://encrypted-tbn1.gstatic.com/images?q=tbn:ANd9GcSnbOHxcTg1zKwPisaX7CsIIbvk6IXj7jHqVD-KWbfGfbShVG4uzGp4pg

Anyways, does anyone have some advice either for achieving pose recognition or for unlocking the opencv issue?

Thanks!

I believe I´m leaning to use http://www.nickgillian.com/wiki/pmwiki.php/GRT/GettingStarted

Discarding the Haar finder.

Just tested GRT (https://github.com/nickgillian/grt) for a couple of days, using the ANBC algorithm for detecting static skeleton poses and works great.

I tested it on VS2012 with OF 084.

Cheers.