I'm back.

Good news is, I have a working draft.

Bad news is, it's not working.

The problem lies in both the imprecision of the accelerometer and the sampling frequency/technique, IMO. The problem being that if you start recording, then you slide the phone forward, then backwards, the numbers don't add up.

In theory, the math is simple. 1G is 9.8 M/s/s, so a 0.1G acceleration yields a 3.6*9.8 M/s if sustained FOR ONE SECOND. Problem being, I have no idea what the mean acceleration was for that second because I have 10 samples that vary wildly.

In theory, going forward one second at 1G gives you 10 samples with +1000, and going backwards gives you 10 samples with -1000 in the XML. In practice, the don't add up.

In the screenshot I posted, I drive first and second gear, up to some 70-80 KPH, then brake to a standstill, then a long stoplight, seen as the uneventful line in the middle. Problem is, the sum of all acceleration, even corrected by integration, shows that I'm still doing 30, because the braking is not as violent and low acceleration isn't as fine-grained or as important to the equation as high acceleration.

One could also see what happens when standing still. Namely, that there's quite some noise from the accelerometers, and even that noise doesn't add up. That means that over the 30 seconds I stood still the error adds up to AT LEAST 0.5 M/s. This means that the compound error is never going to add up or make sense over a long period of time.

Further compounding the problem is the fact that the car actively leans backwards when accelerating hard, some of the acceleration being spread out between axes, with integration necessary.

I'm still working on it.

I think I need a trial run to calibrate errors, the one I have is a standard drive around, with curves and stuff. And I'm half snowed in.

Volunteers?