interesting drift in Z direction
Opened this issue · 5 comments
Sorry to bother u,,, I got an interesting scene when i test the app_tracking program by pushing a small cart in a flat hallway of my school building..., i get the pose(px,py,pz&qw,qx,qy,qz) result by UDP transportation , and change it to ros-path type, and show it in RViz...The result is interesting that the position-z changes a lot(growing) with going forward ..Previous test in a playground also has the scene, but i think it is the reason that playground is not flat ,imu performance would get worse with the body shake..More interesting is that pz grows with going forward, and decrease with going backward(ahhh...seems like correct, seems like the origin is where zero point,,,Maybe it is because the different gravity parameter between Beijing and Shanghai )...I got some pics showing below..How to avoid this problem?,,,Thank u for your explaination!!
Thanks for you feedback. The z-drift is due to the incorrect IMU-camera calibration. However, it can be tricky to perform the IMU extrinsic calibration, so we simply assume it's identity as the factory calibration. Please try Kalibr (need extra steps) to do the extrinsic calibraiton.
ok.thank u...observing from the calib file, seems like I need to change the extrinsic param "D_T_C_r"..I would retry to calibrate it by kalibr
D_T_C_l: CamLeft to Device
D_T_C_r: CamRight to Device
D_T_I: Imu to Device
In our typical setup, we set the CamLeft as Device. {Cl} and {I} should be aligned on the sensor (close to Identity).
report-imucam-homelingdata.pdf
here is my kalibr calibration result,,,the reprojection error is around 0.5pix ,,and seems like the timeshift between cam and imu is very bad(i don't know why..)
Anyway, i change the D_T_I(IMU frame to Device or CamL) from identity to
- [0.99999646, -0.00239824, -0.00115113, 0.01865676]
- [0.00241252, 0.99991809, 0.01256931, 0.0103928]
- [0.00112089, -0.01257204, 0.99992034, -0.03574853]
- [0, 0, 0, 1]
...But, the experiment result seems like nothing useless(very sad ...orz)
thanks very very much for your unselfish help!!!
The kalibr result looks reasonable. It will be helpful to record some data and share w/ us.
You may email us at ext_roboticvision@baidu.com.