0

I have data on the position of the drone from a tracker, as well as estimates of velocity and acceleration based on that data from a Kalman filter variant that seem reasonably good.

I am looking to predict (/give a reasonable guess of) the position of the drone multiple seconds in the future, and I am uncertain if the best tactic is to keep estimating jerk, snap, crackle & pop and include them in the computation, or if there is a better method that I am missing.

Any help appreciated.

TheSandman
  • 9
  • 3

1 Answers1

0

If you have estimated velocity and position, then pass it to your kalman filter. Then pass the output of kalman filter back to it and in this way loop it over again and again. See what happens. I used this method for 2D prediction and my predicted path improved with increasing number of datasets. Also you may need to tweak your control matrix.

Parth
  • 39
  • 10
  • Thanks, I'm looking into it right now. What specifically do you mean by "tweaking the control matrix"? – TheSandman Jun 25 '18 at 17:46
  • Actually I tried my implementation on a number of videos shot by myself. Each of those videos required different values in control matrix. These different values were obtained by hit and trial method. In theory, I studied that control matrix usually is a column matrix of acceleration in x and y directions. For 2D motion these accelerations are constant (ax =0 and ay = 9.81) but for me ax = 0 and ay = 350 worked in one video while ay = 2500 in the other video. I dont know why should someone tweak values like that. – Parth Jun 25 '18 at 17:52