Kacem,
Just noticed this thread. I had spent a couple of days looking at the feasibility of building one of these from a 6-axis mems a year or so ago, I've mentioned it here on scubaboard in response to some threads on "what should we invent for scuba".
My conclusion was that the drift error using available devices I was looking at made it impractical. I think I came up with something like 1/4 mile max error for a 1-hour dive.
It sounds like you're
1) Identifying newer, intrinsically more accurate mems devices.
2) Doing a cleverer job of filtering the noise in s/w.
3) The idea of remotely mounting the sensor itself on the tank where it's less likely to be rotating/moving relative to diver position. I was imagining a hand-held or wrist-mounted unit and resolving the rotational attitude of the device by integrating rotational accelerations, in order to know how to interpret the translational accelerations in real space, causes a lot of the drift error.
I'm mentioning this because I had one idea for minimizing drift error that you haven't mentioned here, that I thought I'd pass on: A button the diver can press that says "Right now I'm not moving and not rotating. Any linear or angular velocities derived from integrating accelerations can now be set to zero.".
Of course, if the diver lies about that the errors are worse, but the potential for reducing error if used correctly seems worthwhile. If used once at the turnaround point, the total drift error at the end of the dive might be halved.
I'd be really interested in one of these in the < $500 range, please keep working on it, and also keep us salivating with occasional progress reports.