Issues with Low Obstacle Persistence and Obstacle Clearing

[kaidegast]

Hi,

I am attempting to use RTAB-Map on a custom mobile platform and have encountered several issues related to sensor configuration and obstacle persistence.

The robot is currently equipped with a 2D LiDAR mounted on top and an Intel RealSense camera mounted lower at the front. The primary issue concerns low obstacles that are visible to the camera but not to the LiDAR. With the current configuration, obstacles detected by either sensor are correctly added to the map; however, when the robot loses visual contact with low obstacles, those obstacles are subsequently cleared from the map. This behavior can be observed in the video below:

rtabmap_test_3.webm

I've tried setting 'Grid/3D':'true' and using the /octomap_grid as the map. Here i have encountered a different issue. Because the robot must pitch forward and backward during acceleration and deceleration, the camera observes vertical motion in the scene. For example, when a person is present, the camera captures points along the full height of the body. After the person leaves the scene, RTAB-Map struggles to correctly clear the points corresponding to the upper body, which results in the obstacle not being fully removed from the map, as illustrated in the image below:

I have shared the relevant rtabmap.db, rosbag file and rtabmap launch file relative to these simulations in the following shared folder:

https://drive.google.com/drive/folders/1y1layPGySpHaXifT6JryBMosfNa8IoRf?usp=sharing

Any guidance or recommendations on how to address this issues would be greatly appreciated.

[matlabbe]

Related posts:

• http://official-rtab-map-forum.206.s1.nabble.com/How-does-rtabmap-deal-with-dynamic-objects-tp4735p8255.html
• rtabmap accuracy degradation in dynamic environment rtabmap_ros#852
• Map not updating the moving object rtabmap_ros#159 (comment)
• OctoMap fails to fully detect moving objects rtabmap_ros#1194
• http://official-rtab-map-forum.206.s1.nabble.com/How-to-avoid-detecting-obstacles-passing-in-front-of-a-running-robot-tp10040p10063.html
• http://official-rtab-map-forum.206.s1.nabble.com/Obstacle-detection-for-RTABMAP-tp5954p5960.html

Dealing with with dynamic obstacles in 3D can be difficult at the mapping level, if they cannot be filtered before feeding data to rtabmap.

Based on comments in the referred links above, if you don't need to track 3D obstacles in the static map, I would use Grid/3D=false with ray tracing on (Grid/RayTracing=true). Then for your specific case of tracking a 3D obstacle that was seen by the camera but not the 2D lidar, and the camera moved in a way that the obstacle (still there) is not in the camera's FOV, I'll suggest to handle this at the local costmap level with a VoxelGrid plugin. An example of config can be seen here for the turtlebot3 RGB-D + Scan demo. Some dynamic obstacles (walking people while the robot is moving) maybe stay in the local costmap, but they would be eventually cleared after the robot leaves that area, which would not be the case for the static map.