I've been working on upper body skeletal tracking with Microsoft Kinect using OpenCV since 5 days now... See the picture. I implemented the algorithm presented in this paper. To let people who are interested in it know, it's not perfect.
It lacks consistency, requires a new computer to be responsive, and can't recognize the arms when they are in front of the body.
Still, hats off to HP people, because what they published is still way better than people who fake skeleton tracking by only doing a simple distance transform and adaptive thresholding, claiming that they do skeleton tracking.
Would you like to know what is perfect?
I just received an email saying that PrimeSense (basically the company that made Microsoft Kinect) already decided to release the alpha version of OpenNI.
I will try it out tomorrow with the Kinect and see how it goes.
I feel defeated and frustrated but very happy at the same time. OpenNI will finally make the interaction work as it should.
Download OpenNI here and PrimeSense Sensor very first version here. See the perfect real full body skeleton tracking with OpenNI here.
