I am working on a sensor fusion problem in a packaging plant. An item comes down the conveyor and is measured by a number of sensors, the outputs of which I am looking to fuse together to improve my state estimation. The sensors measure a number of parameters of the item, such as density and colour. I also have prior knowledge about the item, which I am treating as just another sensor with associated uncertainty.
This question Kalman Filter to improve sensor reading without prediction addresses probably 75% of my question, that it is indeed possible to use Kalman filters in a 'stationary' manner. But I have a couple of follow-ups.
Is this a good method for solving my problem, or would another method like maximum likelihood estimation be better? If so, what are the differences between the two? My approach to implementing the Kalman filter will be to compare the two distributions in the update step, with no change happening in the predict step. Does this reduce in some way to maximum likelihood estimation?
Some of the parameters of the item are correlated with each other, albeit not perfectly. How would I go about incorporating this into my prediction? I want to do something like this: estimate parameter A, estimate parameter B, use the imperfect relationship between the two parameters to improve the estimation. Do I need to do more update steps here or would this be something that would be handled by the multidimensional implentation of a Kalman filter?