First experiments, tests and small results
The last two days I installed DWARF on my PC here:
- Linux: Fedora (Redhat). Why? There was only the choise between Mandrake and Fedora, so I tried Fedora! And it works!
- Why DWARF? I want to develop a distributed collaborative application so DWARF seemed to be right for me!
- TODO: Really LEARN DWARF. But that will come with time!
Combining Inertial and vision
- There are two reasons why you want to combine inertial and vision based tracking:
- more robust and stable tracking, where you try to decrease the flickering (vision) and the drift (inertial)
- predict the movement of the camera. One big problem is, that caused by computation time and latency virtual objects placed in the real world often seem to lag behind the reality. There are several approaches for solving this problem:
- (Extended-)Kalman-Filter: mostly used. It linearises the input with Taylor approximation. But that can sometimes become a problem. It is also computational very expensive.
- sequential Monte Carlo methods
- particle filters [gordon97hybrid]
- First I wrote some classes to encapsulate the ARToolKit? API to "nice" OO. Especially it's now possible to start the marker detection and the display loop in seperate threads.
- So I was potentiated to do several tests. I hab one EyeToy? Camera (D-Link, USB) and one FireFly? (Firewire)
| Test || Resolution || EyeToy || FireFly |
| || || fps || CPU usage || fps || CPU usage |
| only marker detection || 640x480 || 17 - 25 || 40 + 30 || 30 || 50-85% |
| only drawing the captured video image || 640x480 || 6-7 || 60 + 15 || 14 || 95-99% |
| both combined (drawing/marker detection) || 640x480 || 5-8/7-12 || 60 + 25 || 7-9 /10-14 || 80-95% |
| only marker detection || 320x240 || 24-25 || 10-30 +10 || 30 || 22-50% |
| only drawing the captured video image || 320x240 || 24-25 || 30 + 10 || 20 || 95-99% |
| both combined (drawing/marker detection) || 320x240 || 8-14/13-17 || 30- 50 || 10-12 /13-15 || 60-95% |
- I also played aroung with OpenCV? but the frame rate is even worse.
Natural Feature Tracking
I tried to get the Natural Feature Tracking of Kato working but with the EyeToy?
Camera it was almost impossible! Today I had a try with the FireFly?
! It worked better but still not very stable.
I thought of perhaps improving the performance of tracking problems with CG (C for Graphics). CG enables you to comfortably program your GPU which is in several issues multiple times faster than the CPUs.
- 23 Sep 2004