Relative Camera Pose Estimation from essential matrix

1
votes

I want to use a single camera to estimation relative Camera Pose by frames.

For achieving this, following steps are adopted.

  1. calibrated camera get the camera parameter K
  2. Finding point corresponding using SIFT/SURF
  3. Fundamental Matrix Identification ;using the findFundamentalMat（），and input the matches points 2D
  4. Estimation of Essential Matrix by E = K'FK and modifying E for singularity constraint
  5. Decomposition Essential Matrix to get the rotation, R = UWVt / UW'Vt (U and Vt are obtained SVD of E) ,and get the translation t;

But after doing that,the translation t is too small.assume the initial camera position is(0,0,0).After camera moving forward 20cm,the next pose is(0,0,20),the real translation should be 20cm .The level of the two is quite different.

what's wrong? how to obtaining the real translation from translation matrix?Could you help me?

1
opencvcameracomputer-vision
are your camera calibration parameters scaled the way you assume it? If your new pose is 20 and you assume it to be 20cm, might it just be scaled in cm? What kind of value would you expect? - Micka
@Micka Thank you ,the camera calibration parameters are in pixel。And actually the camer move forward 20cm .so I think the translation should be (0,0,20);The camera is monocular.I want to get the real translation in CM. - user4705465
@Micka Thank you ,the camera calibration parameters are in pixel。And actually the camer move forward 20cm .so I think the translation should be (0,0,20);The camera is monocular.I want to get the real translation in CM. - user4705465
afaik in camera calibration parameter there is encoded how big one pixel is on the sensor (= in real world). How do you perform calibration? - Micka
ah ok, your measured translation "should be" (0,0,20) but it isn't? So what does your algorithm measure instead? - Micka

1 Answers

2
votes

Essential matrix is only defined up to scale. The translation vector you get from decomposing E is a unit vector. You cannot get the actual distance between the cameras in world units without more information, such as a reference object of known size in the scene.