Ask Your Question
0

Multiple sensors using robot localization

asked 2019-05-28 12:58:34 -0500

tiagojdias gravatar image

Hello, currently i'm using an EKF from the robot localization package to generate the odom->base_link transformation. For this, i'm using three sources: wheels velocities, laser and visual. Since visual and laser approaches rely heavily on features my approach was to derive instantaneous velocities from two consecutive measurements in time (yea i know they can be noisy).

In this EKF i'm only using velocities so i don't have any reference for position, thus it seems like that my filter covariances for a two d mode starts to grow slowly for position (velocities are fine).

Besides all this, i never have jumps or nothing on my global EKF which takes this velocities from the first ekf and fuse them with a global estimator such as AMCL.

My questions are:

  • can the parameter dynamic process covariance help me in this case? Or
  • since i'm not relying on the state prediction (not feeding any dynamics to the my prediction stage) is generating this behaviors?
  • Or is just some issue with the integration from velocities to position?

I'm still very new on using and tuning Kalman Filters ;)

Thank you in advance.

edit retag flag offensive close merge delete

Comments

I'm not clear what the actual issue is. Are you trying to get the state estimate to jump/correct when you fuse it with AMCL data?

In any case, I can't comment at all unless you provide your full EKF configurations, as well as a sample sensor message for each sensor input.

I will say that you should not be fusing the output of one EKF into the second. Just fuse the same input sources in both.

Tom Moore gravatar imageTom Moore ( 2019-06-06 02:25:20 -0500 )edit

Any update on this? If not, may I close the question?

Tom Moore gravatar imageTom Moore ( 2019-08-22 06:28:21 -0500 )edit

1 Answer

Sort by ยป oldest newest most voted
0

answered 2019-08-25 11:33:37 -0500

tiagojdias gravatar image

You can close the issue.

I was able to solve the problem after some tune on each sensor covariance.

edit flag offensive delete link more

Your Answer

Please start posting anonymously - your entry will be published after you log in or create a new account.

Add Answer

Question Tools

2 followers

Stats

Asked: 2019-05-28 12:58:34 -0500

Seen: 198 times

Last updated: May 28