Improving RTABMap performance

[MidKnight]

Hello @matlabbe, greetings to you!

The aim of my project is to use rtabmap_ros2 (humble) to implement 3D mapping and using the same for autonomous navigation of a drone (PX4 and mavros). I am using intel RealSense D435 camera as my RGBD sensor. I tried implementing RTABMap from the official github repository with docker, and all default parameters. I am facing the following probelms: -

1. Frequent odometry loss.
  - The odometry gets lost, and the map construction gets stuck at the frame where odom was lost.
  - I tried using the rest odom service which allows the map to be constructed from the point where the odom gets lost and I reset it, but the map in that case remains incomplete since the previously constructed map gets lost.
  - I believe that upon revisiting locations, the lost map can be recovered, is that correct?

2. No loop closures.
  - I did not see any loop closures happening, while revisiting locations.

Can you please suggest any resources I can refer for solving these problems?
Can you please suggest any parameter values to be set for achieving correct performance?

This is my workstation's configuration.



This is the command I use:

xhost +local:root 
sudo docker run -it --rm \
 --user $UID \
 -e ROS_HOME=/tmp/.ros \
 -e OMP_WAIT_POLICY=passive \
 -e ROS_DOMAIN_ID=55 \
 -e DISPLAY=$DISPLAY \
 -v /tmp/.X11-unix:/tmp/.X11-unix \
 --network=host \
 --ipc=host \
 --runtime=nvidia \
 -v ~/.ros:/tmp/.ros \
 introlab3it/rtabmap_ros:humble-latest \
 ros2 launch rtabmap_launch rtabmap.launch.py rtabmap_viz:=true database_path:=/tmp/.ros/rtabmap.db rtabmap_args:="--delete_db_on_start" \
    depth_topic:=/camera/camera/aligned_depth_to_color/image_raw \
    rgb_topic:=/camera/camera/color/image_raw \
    camera_info_topic:=/camera/camera/color/camera_info \
    approx_sync:=false \

I would appreciate any help from you, Thanks in advance!

Regards,
MidKnight

[matlabbe]

Since you are using D435, the biggest upgrade you can first do is not use the RGB camera. Use the IR cameras in stereo mode or in Mono+Depth mode. See https://github.com/introlab/rtabmap_ros/blob/ros2/rtabmap_examples/launch/realsense_d435i_infra.launch.py for example. The reason is that you get perfect time sync between Depth and Mono left streams, less motion blur and larger FOV.

After that, it depends on the frame rate you can process frames. If frame rate is lower than 10 Hz, the odometry would likely to get lost more often.

cheers,
Mathieu