I'm interested in this issue, because a friend is a caver, and complains of the difficulties of surveying with slackness amongst the data collectors (cold, wet, muddy, impatient) I envision a device about the size of a rugby football, padded, waterproof, with few controls: start, "still" (pause every ten minutes or so with the device placed on the floor (as during a crawl), or when you quit for the "night") and "stop", recording temperature, atmospheric pressure, and magnetic field. Precision and cost are of course opposites.
Anyway, granted that you have the accelerometers and gyroscopes for two lots of three measurements, and a rapid scan of these parameters (time must be accurate!), in my ignorance I don't see what the problems are with interpreting the results. I would think that you'd calibrate the device for motion first by starting on a table (properly flat and level) and smoothly moving the device one metre (or yard, etc) in a straight line - say the x-direction and back, and again the y-direction. The z-direction is affected by gravity and providing the sensors are correctly aligned, you can thereby calibrate that. Then rotate the device carefully (interchanging xyz) and repeat.
The resulting data record would be a sequence of integers in each direction, say the x-direction. You know that there was an acceleration in the x-direction, and then a negative acceleration. It started from v=0, attained some maximum v and then returned v to zero. You thus should be able to detemine the zero acceleration integer value (as when the device it at rest) and from the integral and known durations, determine the scale factor to produce an x displacement in the chosen unit (of metres, yards, feet)
Depending on accuracy and the table (or floor) size, greater distances should be tried.
Theorising further, there should be no need to align the movements with the physical sensor orientation, indeed I'd be tempted to consider having say three sets of sensors, each with their own orientation (and not orthogonal multiples) with a view to averaging the measurement, and for this purpose, three sets of sensors each of exactly the same orientation should each give the same numbers which is not what is wanted. Because a precision of one part in a thousand (0 to 1023) which is really +-500 is alas not so high, and as time wanders on, error grows (as sqrt(time) at a guess if you're in luck) Thus, whatever the orientation of the sensors in the device, your calibration would be with respect to the flat back of the device as defining the xy plane. Thus, with an arbitrary orientation of sensor, the initial state of motionless, flat on the table, will show where down is in the sensor orientation, and the movements in x and y and z will allow you to find the conversion to the device xy (its flat back) and z (perp. to its back)
For field use, your device must be able to determine the direction of "down" so as to remove the standard acceleration of gravity. Thus the gyroscope readings, to determine change in attitude of the device (and thus the direction of down) so as to remove the unwanted acceleration so that the acceleration that remains applies to your change in position. The gyroscope orientation however does not rotate with the earth, so there are further details but I've typed enough already...