RTABMAP and human in the scene

[Paco]

Hi, I'm looking for information about how RTABMAP behave with mobile human in the scene ?

- Should the robot create a map BEFORE any use with human in the environment ?
- Can we use real time mapping and navigation with human presence ? (Any filtering option ?)

I'm wondering if any work exist in this category for SLAM with RTABMAP to allow navigation with human, if it is already possible to filter human, and if not where should I look first ?

IROS 2014 example (https://www.youtube.com/watch?v=_qiLAWp7AqQ) seems to manage human filtering (the map generated doesn't contain any human) but I didn't find any information about this in the wiki, doc or forum.

Thanks !

[matlabbe]

Hi,

There is no explicit filtering of humans. Dynamic objects would be cleared in the grid map if the camera is still watching the same place after the object has moved. In the IROS video, there was no people during the mapping phase.

The options are:
1) Map the environment without people, then go for localization mode
2) When doing SLAM, when the robot is not moving, it will not add new stuff to map (won't clear either). So a way to not add humans during SLAM is to stop the robot when someone is approaching until he passed the robot. While moving, as stated above, it will depend if the camera is seeing the same area again without the human.

Note that only the 2D occupancy grid maps and 3D Octomap can clear moving objects. The cloud_map output (or MapCloud rviz plugin) keeps/shows everything.

cheers,
Mathieu

[Paco]

Hi Mathieu,
Thanks for the details !

But we can see people on the background of the video, they are too far from the depth sensor so they are ignored ?

https://www.youtube.com/watch?v=_qiLAWp7AqQ

With some clustering algorithm like K-mean or M. Munaro et al. work (link) we cloud filter the point cloud before rtabmap process. This would be possible isn't it ?

[matlabbe]

Hi,

In the video, the people are out of range of the point cloud data, so they are not added. The approach you linked is interesting indeed, and could effectively be done before the mapping process: either regenerate the depth image without people, or use the output cloud with the scan_cloud input topic of rtabmap and tell rtabmap to use scan "Grid/FromDepth=false" to create the grid map.

cheers,
Mathieu