Realsense R200 with rtabmap rgbd odometry

edit

Hi,

You can read the Odometry Lost section of this tutorial: https://github.com/introlab/rtabmap/w... .

There seems to be a lack of features. Some parameters that you can tune:

<param name="Odom/MaxDepth"            type="string" value="0"/>
<param name="Odom/InlierDistance"      type="string" value="0.1"/> <!-- when EstimationType=0 -->
<param name="Odom/EstimationType"      type="string" value="0"/> <!-- 0=3D->3D, 1=3D->2D (PnP) -->

Odom/EstimationType=1 generates more inliers, though not necessary more precise odometry.

If you can record a small rosbag (like 10 sec), I could check it more deeply:

$ roslaunch realsense realsense_r200_launch.launch
$ rosbag record camera/depth/image_raw/compressedDepth camera/color/image_raw/compressed  camera/color/camera_info

UPDATE I tried your rosbag like this:

$ roscore
$ rosparam set use_sim_time true
$ rosrun tf static_transform_publisher 0 0 0 -1.5707963 0 -1.5707963 base_link camera_color_optical_frame 100
$ rosrun tf static_transform_publisher 0 0 0 0 0 0 camera_color_optical_frame camera_depth_optical_frame 100
$ roslaunch rtabmap_ros rgbd_mapping.launch rgb_topic:=/camera/color/image_raw depth_registered_topic:=/camera/depth/image_raw camera_info_topic:=/camera/color/camera_info compressed:=true frame_id:=base_link rtabmap_args:="--delete_db_on_start" estimation:=1
$ rosbag play --clock 2015-10-20-22-55-08.bag

Notes:

• While the depth is registered to color, their frame_id could be the same (here I put null transform between color and depth frames).
• On live experiments with the sensor, you can remove compressed:=true.

Result:

I have similar results with estimation:=0 too. estimation argument of rgbd_mapping.launch is in reality the Odom/EstimationType parameter of the odometry node:

$ rosrun rtabmap_ros rgbd_odometry --params | grep Odom/EstimationType
Param: Odom/EstimationType = "0" [Motion estimation approach: 0:3D->3D, 1:3D->2D (PnP)]

• Odom/EstimationType=0: Transformation estimation is done with 3D feature correspondences, which minimizes the euclidean distance between the inliers. Parameter Odom/InlierDistance=0.1 can be tuned to include more inliers or to be more precise.
• Odom/EstimationType=1: Transformation estimation is done with 3D->2D feature correspondences, which minimizes the reprojection error on the next frame. Parameter Odom/PnPReprojError=5.0 can be used to include more inliers or to be more precise. The advantage of PnP is that the depth values for the next image are not required for the estimation, thus if there are features which don't have depth values, they will be still used. In the previous approach, they cannot be used as the depth is required.

cheers