stereo libviso2 doesn't seem compute an accuracy odometry

asked 2012-09-20 06:50:35 -0500

aldo85ita
252 ●17 ●18 ●31

updated 2012-09-26 22:16:40 -0500

Hi everybody, I've tested viso2_ros wrapping ad it seems to have a low accuracy computing odometry. I run demo contained in libviso2 library (I mean the original C++ library http://www.cvlibs.net/software/libviso2.html ) using the set of images raccomanded by the author. When I run the demo: ./viso2 ../2010_03_09_drive_0019 I get computed camera pose for each processed couple of images (in stero mode), and the latest show me the following matrix pose:

Processing: Frame: 372, Matches: 331, Inliers: 50.1511 %, Current pose: 
   0.7009269    0.0626044   -0.7104803    1.8854907 
  -0.0800644    0.9967505    0.0088415  -15.5700499 
   0.7087250    0.0506869    0.7036616  134.6691852 
   0.0000000    0.0000000    0.0000000    1.0000000

As you can see Y direction that mach with the direction from the bottom to the top, is increased up to -15 meters. It means that the car has flied (or the accumulated odometry error on Y axis is grown a lot). Is it normal? Can I improve it? Consider that this case use stereo visual odometry that is less accurated than visual mono odometry.

edit retag flag offensive close merge delete

add a comment

1

answered 2012-10-02 00:09:06 -0500

aldo85ita
252 ●17 ●18 ●31

The solution is to set the camera pitch and camera height both in /base_link -> /camera tranformation and in "_camera_height:=1.57" and "_camera_pitch:=-0.08" properties:

To test:

export GSCAM_CONFIG="multifilesrc location=\"/home/aldo/Documents/visual_odometry/libviso2/2010_03_09_drive_0019/I1_%06d.png\" index=0 num-buffers=-1  caps=\"image/png,framerate=\(fraction\)19/10\" ! videorate framerate=19/10 ! pngdec ! ffmpegcolorspace"
echo GSCAM_CONFIG= $GSCAM_CONFIG
roscd gscam
cd bin
rosrun gscam gscam &
rosrun image_proc image_proc /image_raw:=gscam/image_raw /camera_info:=gscam/camera_info &

roscd tf
./bin/static_transform_blisher 0 0 1 -1.57079633 0.0 -1.65079633 /base_link /camera 50 &


roscd viso2_ros
rosrun vi2_ros mono_odometer _camera_height:=1.57 _camera_pitch:=-0.08 _ransac_iters:=2000 _inlier_threshold:=0.00001 _motion_threshold:=100.0 /image:=image_rect_color /camera_info:=gscam/camera_info

edit flag offensive delete link

Comments

Yes, when you're using the mono vo you have to set the mono vo specific parameters (i.e. height and pitch) as well obviously. Your question was just about the stereo vo though. I'm glad you were able to resolve your issue.

ThomasK ( 2012-10-02 11:07:15 -0500 )edit

Yes, @ThomasK, you're right: I tried stereo vo because I didn't be satisfied of mono ov, but for both of them I set wrong parameters: in stereo I missed camera pitch and in mono vo I set camera pitch and height only one time each one instead of twice. Thank you very much for your help.

aldo85ita ( 2012-10-02 11:26:49 -0500 )edit

why do you need camera pitch for stereo visual odometry? Can you tell me how you solved this problem? I also met with the problem of an inaccurate y estimation.

xgdong ( 2015-04-04 22:47:03 -0500 )edit

add a comment

3

answered 2012-09-26 08:45:41 -0500

ThomasK
360 ●8 ●11 ●17

libviso2 works perfectly fine when used with correct parameters. The problem you're seeing is probably caused by your initial transformation being different from the initial camera pose in the KIT dataset. As far as I remember the KIT camera is slightly facing downwards, hence the camera is recording a constant upward motion (relative to the camera) so when your initial camera position is facing forward instead of slightly downward like the KIT camera your odometry will end up going upwards. The calculated odometry is still correct but just rotated by the initial offset between your camera and the dataset's camera.

You can look up the sensor calibration on KIT's dataset website.

edit flag offensive delete link

Comments

Using libviso2, It's possible to set camera pitch with mono-odometer case (in VisualOdometry::parameters struct). How can i set it with stereo case (the showed example above)?

aldo85ita ( 2012-09-27 02:17:59 -0500 )edit

Well I'm assuming you're using libviso2 with ROS and have some TF frame set up that's being used by your camera? libviso2 just calculates the odometry as perceived by the camera, so you need to make sure that the tf frame you're using for your camera has the correct orientation relative to the map.

ThomasK ( 2012-09-27 07:37:59 -0500 )edit

Yes I'm using libviso2 package and viso2_ros wrapper. But I also test demo with standard C library with the same result. As you suggest, I'll set base_link --> camera trasformation using static_transform_publisher command and I will update you.

aldo85ita ( 2012-09-27 22:10:50 -0500 )edit

Thank you @ThomasK, after several tests, I discover that is mandatory to set the camera pitch both in "_camera_pitch:=-0.08" of mono-odometer and in base_link-> camera tranform. Otherwise the mono odometry algorithm doesn't work fine. I'll add all settings in the post below.

aldo85ita ( 2012-10-02 00:04:33 -0500 )edit

add a comment

-2

answered 2013-03-11 02:34:37 -0500