updated 04:30 pm EDT, Tue August 9, 2011
Kinect paper outlines computing algorithm
A new paper (pdf) from Microsoft Research has given developers a fresh look at how the Kinect motion-sensing peripheral works. It focuses specifically on the human body tracking component. The algorithm runs on the Xbox console and the Kinect drivers, which is what is provided to games developers.
It is said to provide an ultra-reliable 3D joint orientation and location map which maps a person's pose in Cartesian space. Developers can then build gesture and action support on top of this as a way of allowing gamers to interact with the environment. The paper, due for publishing at an upcoming conference, uses a collection of decision trees (dubbed a decision forest) that's trained using thousands of sample datasets over a larger number of hours of cluster compute hours. This eliminates the need for a calibration before a pose, won't cause problems if the pose is much different in a short amount of time and is more efficient for the visual data from the Kinect camera.
The algorithm is said to be very smooth, reliable and efficient by developers. The x, y and z manipulations can easily be handed off to the GPU for quicker computation and allow for up to 200 frames per second on fairly standard hardware. [via DeveloperFusion]