Don't know anything formal in this field but this is a really great explainer. It's worth mentioning that if devices are miscalibrated such that the distribution is not centered around ground truth, this approach breaks down. Maybe there are more sophisticated ways to solve this that work well with the approach described. Not sure, but curious.
That doesn't pose an issue: the "observation model" of a Kalman filer allows for affine transformations (e.g. scale, rotation, and offset) between the sensor data and the "true" latent state that you wish to estimate. Even better, it can appropriately handle correlational structure between your sensors.
What the Kalman filter equations don't tell you, is how to estimate the parameters of such an observation model. You either have to write it down from first principles, estimate it from "fully observed" data (if you have such a luxury), estimate it (up to a rotation) using expectation maximization (EM), or guess.
