Best approach to fuse odometry?

[jorgemia]

I've been reading some of the forums and it seems some people fuse odometry before feeding it to rtab-map whilst others combine the visual-odometry created by rtab-map with the IMU data.

If I have an IMU, wheel odometry and an intel D435i camera, what would be the best approach to combine all these different sources of odometry for outdoor navigation (in ROS)?

Use robot_localization EKF before and feed rtab-map a filtered odom topic (IMU + wheels)? Would it be useful to generate visual odom after? Can I fuse that as well?

[matlabbe]

There are multiple combinations possible, I would try these two:

wheel odometry -----------------------\
                                       \
IMU -----------------------------> robot_localization -------> rtabmap
                                       /
D435 images ---> visual odometry -----/
                   /
D435 imu ---------/

wheel odometry ------------------\
                                  \
IMU --------------------------> robot_localization -------> rtabmap
                                   /
D435 images---->visual odometry --/

[jorgemia]

Thanks @matlabbe. I actually don't have a second IMU anymore and instead have a T265 camera.

Do you think this approach would work? Would it be overkill?

wheel odometry --------------------------\
                                                           \
T265 visual odom ----------------> robot_localization ----->rtabmap
                                                           /
D435 images ---> visual odometry -----/
                            /
D435 imu ---------/

[matlabbe]

Hi,

I would go with only one visual odometry:

wheel odometry --------------------------\
                                                           \
T265 visual odom ----------------> robot_localization ----->rtabmap
                                                                                      /
D435 images ------------------------------------------------>/

[jorgemia]

Thanks for your answers. Final question...With two zed2 cameras on a skid steer robot facing opposite each other, of the options below, do you think one is better than the other?

zedA odom -----|

zedB odom -----|
                       
wheel odom --- |------> robot_localization --> rtabmap

zedA IMU ------|

zedB IMU ------|

(The odom provided by the zed wrapper node is visual-intertial)

or

zedA ----\
              ----> rtab visual odometry---\
zedB ----/                                           \
                                                           \
wheel odom ---------------------------> robot_localization ----> rtabmap
                                                            /
imu -----------------------------------------/

[matlabbe]

The advantage of the second approach is that both cameras are used for odometry, so better field of view. It won't be necessarily more accurate, though more robust to occlusions.

With the first approach, you should make sure you can transform zed odometry in the base frame of the robto, the frame_id used by wheel odometry. With rtabmap, we can do that.