misalignment from tango multi session
Locked
123
123
Re: misalignment from tango multi session
 
|
Hi mathieu
actually, tango is scotch on top of the lidar(as in design02.jpg).Lidar point source is 4 cm backward Tango IR source (assuming Tango IR is the center of Tango device in RTAB, I center it on lidar). Not the best settup but I'm designing 2 3d print stand to mount the phone properly.   Main goal of the lidar is to catch what tango don't see. With all my test with tango,I known I'm able to scan a large environment, as long as I have good "unique scene" across the environment to make good loop closure. And results are unbelievable. The only problem is the field of capture. I have to make a lot of back and forth between 2 "unique scene" to scan each angle of the environment. According to you, which design is the best? I made 2 other db. One speed, one slow.$ rosrun tf static_transform_publisher 0 0 0.04 1.57079632679 3.14159265359 0 device Lidar 100 https://wetransfer.com/downloads/419a892f1327737fc9a89f0ebf0cc64f20190228143005/435b51e52dde213191db420754db93fc20190228143005/ef7c88 If I move very slowly, lidar look effectively less off-sync from Tango. High speed increase computation time and error on rotation. Don't know if there is a link. |
Re: misalignment from tango multi session
Administrator
|
Re: misalignment from tango multi session
|
|
Re: misalignment from tango multi session
|
Administrator
|
Re: misalignment from tango multi session
|
|
Re: misalignment from tango multi session
|
Administrator
|
Hi,
I did an error, the second camera_info was indeed /tango/point_cloud topic instead, I corrected the post above. Your linked bag is missing the images : $ rosbag info 2019-03-18-15-50-45.bag
path: 2019-03-18-15-50-45.bag
version: 2.0
duration: 28.3s
start: Mar 18 2019 10:50:45.59 (1552920645.59)
end: Mar 18 2019 10:51:13.89 (1552920673.89)
size: 395.3 MB
messages: 6476
compression: none [287/287 chunks]
types: sensor_msgs/CameraInfo [c9a58c1b0b154e0e6da7578cb991d214]
sensor_msgs/Image [060021388200f6f0f447d0fcd9c64743]
sensor_msgs/PointCloud2 [1158d486dd51d683ce2f1be655c3c181]
tf2_msgs/TFMessage [94810edda583a504dfda3829e70d7eec]
topics: /Lidar/points 285 msgs : sensor_msgs/PointCloud2
/tango/camera/color_1/camera_info 102 msgs : sensor_msgs/CameraInfo
/tango/camera/color_1/image_raw 1 msg : sensor_msgs/Image
/tango/point_cloud 33 msgs : sensor_msgs/PointCloud2
/tf 6055 msgs : tf2_msgs/TFMessage (3 connections)However, I checked with RVIZ the data and the TF between device and Lidar is good (fixed). Here is an example just accumulating the lidar point clouds from that bag: https://youtu.be/OoP1yWY9TJQ  The cloud seems good! Example saving all scans in rtabmap: $ roscore $ rosparam set use_sim_time true $ rosrun rtabmap_ros rtabmap -d \ --Rtabmap/DetectionRate 0 \ _odom_frame_id:=start_of_service \ _frame_id:=device \ _subscribe_scan_cloud:=true \ scan_cloud:=assembled_cloud \ _subscribe_depth:=false \ _subscribe_rgb:=false $ rosrun nodelet nodelet standalone rtabmap_ros/point_cloud_assembler \ cloud:=/Lidar/points \ _fixed_frame_id:=start_of_service \ _max_clouds:=10 $ rosbag play --clock 2019-03-18-15-50-45.bag Note that we set Rtabmap/DetectionRate to 0 as we are received the assembled scans at 1 hz. Then exporting the full cloud to PLY (with rtabmap's File->Export Clouds...) and view it in CloudCompare:  There are still some small errors in rotation, but I can think you are on a good way. For the synchnorization of Lidar with TF coming from Tango, the best would be to somewhat synchronize the clock between the phone and computer feeding the lidar, though not sure how to do that. For the 1 second lag on the previous post, I think I have seen it with Tango too, where the clock of Tango gets slowly unsynchronized with remote computer after a couple of minutes, the fix is to reboot the phone. cheers, Mathieu |
Re: misalignment from tango multi session
|
|
This post was updated on .
hi mathieu.
I did some other test, tango points and lidar points are better align. The only things I change is the static transform publisher: from : rosrun tf static_transform_publisher 0 0 .04 1.57079632679 3.14159265359 0 device Lidar 100 to: rosrun tf static_transform_publisher 0 0 .04 1.57079632679 3.14159265359 0 device Lidar 5 and the launch command sequence. my tf tree look like this:  Things look better but that's not perferct. Don't known if it's an general error or only the precision from tango IMU that shift points of lidar when there are far from source. When I show IMU in rviz, it's seems to shake a lot on rotation. Some other things also disturb me. When I launch "rosrun tf tf_monitor", I got this: RESULTS: for all Frames Frames: Frame: Lidar published by unknown_publisher Average Delay: -0.00470797 Max Delay: 0.00404124 Frame: camera_color published by unknown_publisher(static) Average Delay: 0 Max Delay: 0 Frame: camera_depth published by unknown_publisher(static) Average Delay: 0 Max Delay: 0 Frame: camera_fisheye published by unknown_publisher(static) Average Delay: 0 Max Delay: 0 Frame: device published by unknown_publisher Average Delay: -0.158295 Max Delay: 0 Frame: imu published by unknown_publisher(static) Average Delay: 0 Max Delay: 0 Frame: laser published by unknown_publisher(static) Average Delay: 0 Max Delay: 0 Frame: start_of_service published by unknown_publisher(static) Average Delay: 0 Max Delay: 0 All Broadcasters: Node: unknown_publisher 198.213 Hz, Average Delay: -0.158295 Max Delay: 0 Node: unknown_publisher(static) 198.109 Hz, Average Delay: 0 Max Delay: 0 is that normal there are all named "unknown publisher"? and sometimes during scan,this message pop: Warning: TF_OLD_DATA ignoring data from the past for frame start_of_service at time 1.55437e+09 according to authority unknown_publisher Possible reasons are listed at http://wiki.ros.org/tf/Errors%20explained at line 277 in /tmp/binarydeb/ros-kinetic-tf2-0.5.20/src/buffer_core.cpp I tried reset tf from rviz bottom button, but don't see any difference A last thing: In tango ros streamer, I can't acces the setting "enable color camera". Says "requires API >23. Is that why my bag is missing the images? I'm thinking about arduino IMU module to test it's presicion compare to the tango IMU. Do you think that can help? I have doubt about this mapping method. Here we map with the tango IMU precision. That is why I have noise on lidar points (same noise than tango point cloud) while lidar is a very precise device. Can we make the pose estimation in a other way? Is there a way to find features in lidar points to estimate its transform in the 3d world? then check every second if this lidar pose estimation and Tango imu pose are enough close to validate lidar transform and points, to use lidar precision, not IMU precision.Then use images and when RTAB-map find a good loop closure, set Lidar pose estimation and IMU pose at the same location. I read the "RTAB setup with a 3D LiDAR" thread. Is it similar to what I'm thinking? DB from my last test,scan test start at frames 765, shaking IMU at the beginning. https://we.tl/t-NrAWfl2ERG thanks for all your help, results start looking good. cheers. |
Re: misalignment from tango multi session
|
Administrator
|
Hi,
I've found that the covariance set for tango odometry is not set, thus the loop closures add more errors in the map than what they can correct. I updated the tutorial. You need to start rtabmap.launch with (based on what I've set for RTAB-Map Tango): odom_tf_linear_variance:=0.0001 odom_tf_angular_variance:=0.000001 At the beginning, the camera is looking straight up toward the ceiling, where there are not a lot of discriminative features, thus Tango is drifting a lot more. Beware that Tango drifts more when the camera is looking textureless (or with very repetitive textures) areas. Those are the point clouds created with only Tango odometry, see how there are multiple layers on the ceiling.   It is possible to refine the odometry poses with the laser scans. However, in this database when rotating there is not always a lot of overlap with the previous scan. Increase Rtabmap/DetectionRate to record more frames/scans (if you set it to 0, it would record at 5 Hz matching max tango cloud rate). Also, there is seem to still have not a good synchronization between tango and the lidar. Not sure if it is a timestamp problem (stamps not matching between the devices) or some TF delay somewhere. For the static transform publisher, not sure setting at 200 Hz vs 10 Hz is the solution, as static transforms don't change over time. You may be lucky that time the stamps between Tango and host device were more synchronized. Here is a comparison between tango odometry and when refining using scan matching:   In the second image the scans were downsampled for he matching, but we can see that when the scans are matching the tango clouds don't, which means a synchronization/TF problem between lidar and tango. I don't think the arduino IMU would be more accurate than what Tango is using, though you may get a better synchronization with the lidar. With lidar-only + IMU (without tango), you may look at this velodyne example or this ouster example. cheers, Mathieu |
Re: misalignment from tango multi session
|
|
Re: misalignment from tango multi session
|
Administrator
|
Re: misalignment from tango multi session
|
|
Re: misalignment from tango multi session
|
Administrator
|
Re: misalignment from tango multi session
|
|
Re: misalignment from tango multi session
|
Administrator
|
Re: misalignment from tango multi session
|
|
Re: misalignment from tango multi session
|
|
Re: misalignment from tango multi session
|
|
Re: misalignment from tango multi session
|
Administrator
|
Hi Stephane,
Looking at the generated point cloud from laser scans, there are strange overlapping walls (see section inside blue rectangle):  I looked in the database and all links are pretty good. The problem here is the glass wall in the right room. There an example of single scan and corresponding RGB view: see how the glass wall we see in RGB image is causing reflections in the scan (doubling the room behind the glass wall), shown in blue rectange on right top view of the scans (the green triangle is the camera frustum):  This kind of situation (e.g., glass, windows, reflective surfaces) is causing the mirroring room problem as seen in the first image of this post. Ideally, it would be nice to select those points behind the wall in DatabaseViewer and delete them, so they are not added to global map afterwards. However, that feature is only implemented with depth images (like removing mirrors in this tutorial)... Another problem in the database: the TF between camera and lidar seems a little off in "yaw", so the walls in RGB-D image doesn't match the walls in the scan:  EDIT For your question: You may try traversing backward the doors, so that the scans can still be registered with the previous ones. cheers, Mathieu |
Re: misalignment from tango multi session
|
|
Re: misalignment from tango multi session
|
Administrator
|
| Free forum by Nabble | Edit this page |