matlabbe
commented
5 years ago Hi David,
Thanks for sharing the screen recordings and the database. I could try re-run the mapping with different parameters.
Modifying the grid map only works when memory management is disabled and in localization mode:
- Do mapping without memory management (
Rtabmap/TimeThr=0) - Close database and open it in rtabmap-databaseViewer
- File->Export saved 2d map
- Edit grid map file in any image editor
- File->Import 2d map
- Relaunch rtabmap in localization mode (
Mem/IncrementalMemory=false), it should use the manually modified grid map and not update it afterwards
I noticed that processing time for image registrations (see "Neighbor link refining" and "Add loop closure link" timings below) is quite high on the robot. RGBD/LoopClosureReextractFeatures is not necessary in the new versions as raw descriptors are saved by default in nodes, you will save processing time. As the ground is a plane, you could disable normals segmentation Grid/NormalsSegmentation to save >50% time on creating the local grids. As RGBD/OptimizeMaxError is set, Optimizer/Robust should be false.
RGBD/LoopClosureReextractFeatures=false
Grid/NormalsSegmentation=false
Optimizer/Robust=false
If you want to re-use the same database you created with those new parameters, you can regenerate the map like that (make sure you have latest rtabmap version from source to do this):
$ rtabmap-reprocess --Rtabmap/TimeThr 0 --RGBD/LoopClosureReextractFeatures false --Grid/NormalsSegmentation false --Optimizer/Robust false rtabmap_1turn_optimized.db output.db
I also noticed that as RGBD/NeighborLinkRefining=true, the resulting refined trajectory looks strange when the robot rotates on place ( while the resulting assembled cloud looks okay):
It creates always an "arc"-like correction. Normally, if TF between /base_link and /camera_link is perfect, the correction should not create a systematic arc like that. The error is consistent with all rotations in the map, so the problem would be that static TF is not accurate enough. Note that if the camera TF is fixed and that odometry doesn't drift too much, you could disable RGBD/NeighborLinkRefining to save some processing time. Make sure to set realistic covariance for your input odometry in rtabmap node:
<param name="odom_frame_id" value="odom"/>
<param name="odom_tf_linear_variance" value="0.001"/>
<param name="odom_tf_angular_variance" value="0.005"/>For the jumps during navigation, the local planner could be less constraint to follow exactly the path, to move less in zigzag on relocalization. If you could have a second database, we would be able to replay it against the first map in localization mode to better tune the parameters to decrease the jumps.
cheers, Mathieu
molinadavid
The white line is the trajectory during localization. I've found that at some places, localization is not possible as features are only taken on foreground where depth is valid:
Note how the depth image at the bottom overlaps only where objects have changed.
Loop closures are found more often, though not necessary accurate when not enough features with valid depth can be extracted. Compare that image with the one of the first configuration, now far features in the center of images (where there is no depth) can be used for localization.
The localizations are a little more smoother.
Greetings,
I have created a map using rtabmap_ros and processed the database using rtabmap-databaseviewer, after that when I want to navigate that map I find 2 big problems and I would very much appreciate your help solving them.
The first problem I have is that the correction of the robot's pose some times is very big making the robot "jump" on the map long distances, I would like to know how could I minimize this "jumps" to have the robot navigate smoother.
The second problem is actually related to the grid map, as I mentioned on the title I would like to use the globally optimized grid map for navigation, the problem I am facing is that the optimized grid does not remain constant and get "unloaded" and "reconstructed" again some times, but it does not load the full grid globally optimized.
I tried exporting the grid, cleaning it and then re importing it again using the databaseviewer, it works when it first load the database, but then it gets unloaded and the same as before happens.
I also tried publishing the static grid map using map_server, it works for most of the time but there are some areas where the sync is not perfect and there is a risk of the robot thinking that is navigating on a wall or something like that.
in order to help illustrating my issues I will share some screen recordings and the database used for this navigation.
Video showing rtabmap grid unloaded Video showing static map Database used for navigation
Thank you very much in advance for the help you can provide on this.
Best regards, David