SLAM with stereo zed

[user1245]

Hi,
I'm here for some advices on how to improve slam with stereo zed (some parameters).
I'm using rtabmap in stereo mode and the odometry provided directly from zed.

I need a 2D grid map and a good localization for navigation.

This is the map obtained by a Lidar (with another SLAM):
Map with Lidar

This is the map obtained only from zed camera in stereo mode:
Map with Zed

The problem is that the localization is not very accurate and if I continue to make laps, the map gets worse (odometry problem? loop closure?)

This is my rtabmap_stereo.launch.
Thank you.

[matlabbe]

Hi,

Lidar and Visual SLAM approaches behave differently depending on the environments. In indoor environments where there is a lot of geometry, a lidar-based slam would be more stable over time as loop closures can be detected in any direction (if we assume a lidar with >180 deg field of view). See this paper for a comparison with RTAB-Map in lidar-mode or stereo-mode. In particular Table 8.

For the results you have, they are what would be expected from a lidar and a stereo setup. Recopying your images here (Lidar on left and stereo on right):


If you loop many times the same area, you may expect more double walls with a visual approach as less loop closures would be found. In particular in textureless environments (e.g., with mostly white walls), visual SLAM would drift more than lidar SLAM.

Note that with RTAB-Map, you can combine the lidar with the camera, see http://wiki.ros.org/rtabmap_ros/Tutorials/SetupOnYourRobot

cheers,
Mathieu

[willzoe]

I had a similar problem. This is the link to my question: https://answers.ros.org/question/348011/when-using-rtabmap-to-build-a-map-how-to-fill-the-blanks-in-the-map-and-remove-objects-outside-the-map/

If there are only stereo camera and IMU on turtlebot2, is it possible to build a 2D grid map that can be used for navigation?

[matlabbe]

Yes, I answered your question on ros answers.