search

rsasaki0109 / kalman_filter_localization

kalmal filter localization
Other
209 stars 45 forks source link

Validate fusion result #5

Open mengen100 opened 3 years ago

mengen100 commented 3 years ago

Hii, I've one more question: How to validate the accuracy of the fusion result?

rsasaki0109 commented 3 years ago

The accuracy cannot be verified with sample data. You can verify the accuracy of this ekf by using it with an open data set with true values, such as KITTI.

mengen100 commented 3 years ago

Thanks!

tankang666 commented 2 years ago

Excuse me, I don't use GPS, I only use Odom. Why is the output topic messy and the data floating? When GPS and Odom are input at the same time, which is used for observation?

rsasaki0109 commented 2 years ago

Sorry, the debugging of the Odom complex may have been poor. I'll fix it as soon as I figure out the problem.

Sang-Beom-Woo commented 2 years ago

Excuse me, I don't use GPS, I only use Odom. Why is the output topic messy and the data floating? When GPS and Odom are input at the same time, which is used for observation?

As far as I know, any type of odometry (wheel odom, visual odom, visual-inertial odom... etc. ) suffers from drift. Fundamentally, there is no way we can correct them without a bounded correction source (e.g. GPS, wifi Beacon, Place Recognition(in case of SLAM)... etc). In order to make odometry in novel EKF, the code needs an extra code bit for odometry prediction.

Meanwhile, in kalman_filter_localization package, both odom and GPS are used as observation. Odom has a relatively small drift to IMU, the assumption is reasonable for local navigation. However, the pose will keep drifting out from the gound truth as it uses poorly estimated covariance updates.

There might be a wrong explanation, so feel free to make a comment to correct this posting