hi

 Your q is interesting Some conjectures below .hope it helps.

1. Let us imagine  the ref axis is say x-y-z plane  fixed to image

2 Assume image lies on xz plane and camera is centrally placed  on y axis

3 Approx location of camera would be using focal length of camera fc and we have (some known dimensioned object in image) viz lobject/Limag . then ycord=lobject/Limag*fc .Now wrt object ref axis camera is located at(o,ycord,0).

4 Next if camera is also shifted upwards say by z0. then we tilt image around  xaxis and by suitable transformation on image get most acceptable image (aspect ratio?)'it should be possible to work out the angle of inclination and determine the zo value..requires some working

5. Alternatively there probably are codes with suitable transformations where you can just experiment and try.

Cheers

If you search for "camera calibration" (and "extrinsic camera parameters" in particular) you will find a lot of results. Some starting points: http://docs.opencv.org/doc/tutorials/calib3d/camera_calibration/camera_calibration.html and http://wiki.ros.org/camera_calibration.

You need a global reference system before image record. And, it should be fixed a local system in each camera. After a static test you could use tranformation matrices among reference systems.

Thanks, Members. Sorry for the late reply.

I have a problem at hand. "Determining the depth parameter of an object from a stereo pair of images". I have gone through the following steps. 1) Had to extract interest points from both images using Harris detector. 2) Had to find matching points using ICP algorithm. 3) From the matches had to find the fundamental matrix using the norm8point algorithm. 4)  Computed the essential matrix from fundamental matrix and intrinsic parameters of the cameras. 5) Am supposed to find the rotation matrix and translation vector from the essential matrix from with will transform the two point sets and find the disparity to compute the depth. But the Rotation and translation matrices i compute are always wrong. what is the best and efficient method ?

The issue(s) can be any one or more of your steps. Have you considered lens distortion via camera calibration? Have you considered removing possible wrong matches using RANSAC or equivalent? Some of the functions in OpenCV might be of interest: https://docs.opencv.org/2.4/modules/calib3d/doc/camera_calibration_and_3d_reconstruction.html