douglas:
thanks for the lecture
i really do not think a shake reduction system bothers to integrate ot get speed, it probably works more on frequencies, and has a lot more fun with statistics than we would imagine (trying to be "predictive" about it. on how panning myght be detected.. you might be dissapointed: i am pretty sure the key-word here is "thresholds", probably on frequencies, and in the case of canikon, i think it is just "ignore the horizontal component", at least roughly (but i think some newer systems of theirs also have an automatic way of deciding, so i guess they might use thresholds also)
another related hint, which might not be as obvious to most people: integration is actually a very "natural" process, which is "numerized" to be more universal. when you do it digitally, you have the advantage of flexibility, but it is very computationaly expensive. if you know exactly what you want to do, you might be able to design something much cheaper from a computational standpoint (maybe even analogue, in some "extreme" cases), instead of using a generic digital cpu to do it "soft".
imagine this rather rough example: you have a "simple" curve. it is "real", so you need a function to describe it (which is he first step, and headache), and only than you can integrate. or you can just "walk" it with a device, to measure distances on it. integration is in theory capabale of arbitrary precision, given enough computing power, however reality rarely is that precise itslef, and i am sure you know what i mean
, otoh walking it means to have a (simple, but still dedicated) device just for it, which won't be much use for anything else, but sometimes it can work better. otoh, if you start out with "real world", you will always need a device in the real world, a way to digitze it's output (adc) a way to reverse the output from the computing task (dac) and another device to _do_ stuff in the real world, as a result of all this work. as things get complicated, you can save more significant amounts of computing power (which starts to matter in real time applications), but you start spending (exponentialy, perhaps) more on design (and software will always be cheaper, and these days "digital" hardware is hard to beat on price, with all the r&d already pumped into it). when size also matters, things get more complicated. but i am sure you know more about it than i do
(just some food for thought, i am no expert, unlike you, obviously). btw, that's some really cool job you seem to have