I combined opencv tracking (using Camshift algorithm) with OpenGL here:
http://www.youtube.com/watch?v=yEioXZT-lv0
My first recommendation is not to try to access opengl libraries directly - there is a very nice java shell for it called Rajawali
http://www.rozengain.com/blog/2011/08/23/announcing-rajawali-an-opengl-es-2-0-based-3d-framework-for-android/
You can use opengl directly, but I found the interface rather confusing, and Rajawali hides a lot of complexity.
Now for comparing the two-dimensional coordinates of opencv with 3D openGL, an interesting problem is a). you need to worry about how your camera “sees” the 3D world in terms of lens distortion, etc. and b). you lose measurement when you switch from what the camera sees in 2D to a 3D model, and therefore you need to compensate for this loss of information from other sources, such as the size of the tracked object. Vuforia will do this with a reference image, so this is a shame you cannot use.
These issues are discussed in sections 11 and 12 of the O'Reilly book http://www.amazon.co.uk/Learning-OpenCV-Computer-Vision-Library/dp/0596516134/ (although this book covers C / C ++ interfaces and not Java, but usually there is an obvious mapping).
In my example, I did not do anything complicated - the larger the Camshift detection area, the closer I moved the camera position in the 3D model. I used accelerometer sensors to detect clockwise / counterclockwise rotation, etc. Actually, Vuforia would be better for what I was trying to do, but you are here.
The btw model was created in Blender, and you can import it into Rajawali, including basic texturing.
Hope this helps!