Mapping deviates from GPS Odometry

[aswingururaj]

Hi, I am using rtabmap_ros with a ZED camera for a project on driverless car. The mapping deviates drastically from that of the GPS at speeds greater than 10 km/h. I tried changing the parameters that you have listed in issue #102 . But still the map deviates very much after it loses odometry for the first time after launching.What can be possible reasons for this? Is there any other parameters that I need to change to get better results? Aswin

[matlabbe]

Can you record zed data and share it? What is the framerate of the odometry? Which rtabmap version do you use?

[aswingururaj]

https://drive.google.com/file/d/1diho8FUIiswRrAB5QsjozSEl3GN33w7M/view It starts at 60 Hz and slows down as we increase the speed of the car. we are using rtabmap_ros. The red screens don't appear that often but still the mapping is largely deviated from the GPS odometry. Green is GPS and red is rtabmap.

[matlabbe]

You may rotate the red trajectory to better compare with the plan. For the parameters, which rtabmap version are you using? Can you list all parameters you changed? A lot of improvements for odometry have been done in 0.17. Looking at the features extracted, it seems that you are using RGB and depth images of the ZED. With stereo cameras, I recommend to use stereo_odometry instead of rgbd_odometry as far features will be used for motion estimation.

What do you do when odometry gets lost? Do you reset to latest estimated pose?

cheers, Mathieu

[aswingururaj]

We will try to rotate that. We are using 0.11 version. We changed the following parameters: Odom/Strategy=1
Odom/GuessMotion=true Vis/EstimationType=1 Vis/CorType=1 Odom/ResetCountdown=10 We tried using stereo_odometry but there was an error(Didnot receive data for 5 seconds).As of now we are using only the resetcountdown parameter. After the odometry gets lost the plotting is full of errors.Since we are using it in a car it is not possible to come back to the frame at which the odometry was lost. Is there any way by which we can account for the frames that are lost in that time?

[matlabbe]

I suggest to update to 0.17 (binaries are available on Indigo/Kinetic) to get better odometry first. Then for odometry lost, the new version (of rgbd_odometry or stereo_odometry) can publish odometry only if there is a valid velocity with parameter publish_null_when_lost set to false. You can use robot_localization package to integrate the odometry velocity to get a continuous pose estimation like:

<rosparam param="odom0_config">[false, false, false,
                                false, false, false,
                                true, true, true, <!-- Linear velocity -->
                                true, true, true, <!-- Angular velocity -->
                                false, false, false]</rosparam>

When using robot_localization, don't forget to set publish_tf of rgbd_odometry or stereo_odometry at false to avoid TF problems (robot_localization publishes also odom->base_link).

EDIT: robot_localization is a kalman filter, if odometry velocity is not published for a second, it may still interpolate the position.

cheers, Mathieu