1

I am trying to merge point clouds taken from different angles of objects using the Lenovo Phab 2 Pro. I use the framepair COORDINATE_FRAME_START_OF_SERVICEas base and COORDINATE_FRAME_DEVICE as target. I then translate the point clouds using the transformation matrix retrieved with TangoSupport.getMatrixTransformAtTime() with the same timestamp as the point cloud. However the point clouds do not line up as good as I was hoping as you can see in this image:

enter image description here

I assume this is caused by the drifting the device is experiencing while moving around the object. Is there a way to minimize the drifting by using Area Learning or something similar? Or should I try to merge the point clouds in a different way such as ICP Registration?

Björn Englesson
  • 31
  • 7

2 Answers2

1

First I would try to use Area Learning. In a presentation of Sturm(2015) he speaks of an error <1% of the moved distance, see page 23: https://jsturm.de/publications/data/sturm2015_dagstuhl.pdf

Second it seems to be more accurate to collect some features before estimating the pose: Project Tango onPoseAvailable() and getPoseAtTime() discrepancies

I'd suggest to start measuring the scene, moving the device around, so the scale of the visual inertial odometry can be estimated well and after one minute start measuring the depht images. (A button is working great). And transform everything at the end with TangoSupport.getMatrixTransformAtTime().

I think ICP won't work that great because the depth images are rather small and noisy. My accuracy estimation of the developer tablet's depth camera are ~2mm at 1.5 m distance up to ~11 mm at 3 m.

Community
  • 1
  • 1
LewisTherin
  • 11
  • 3
0

In fact, only the fish eye camera and the IMU are use for the localization. Which means the 3D data is not used for the localization. So a calibration error can lead to inaccuracies. You can find the calibration tutorial here.

The drift of the frame-to-keyframe provided visual odometry algorithm can also be responsible of your alignment problem.

In fact, for scanning objects, I would recommend Structure which offers a frame-to-model tracking approach and a higher resolution depth sensor.

Vincent Vidal
  • 323
  • 2
  • 11