I am working on the capture of sports movements, and in particular on X, Y, Z positions of a set of key points. Since measurements are somewhat noisy, I would like to implement a Kalman smoother with a constant acceleration process model, leveraging filterpy.
To my great surprise, it seems like it does not matter what the values of measurement noise and process noise are, as long as the ratio does not change. I can multiply both noises by a thousand, test it with different data, it does not make any difference.
Another question: results are best when my process noise is a 100 times higher than my measurement noise. This seems quite odd, as in most if not all examples I found, the measurement noise is much higher than the process noise.
Any input would be much appreciated, thanks in advance! Here is my code if you need it.
