How to calculate covariance matrix for monocular SLAM?

edit

Hi there, I am using a single camera for visual odometry. The camera is fixed to the robot and I use the common approach of matching descriptors between consecutive frames and then estimating the rotation and translation using SVD. The camera is already calibrated and I already have the points in 3D position.

From the rotation matrix and the translation matrix I can obtain the the rotation angles and the new position x,y,z of the robot.

My question then is: How can I estimate the covariance matrix for theses values in order to publish an nav_msgs::Odometry matrix? I suspect that I have to backpropagate the points from one frame to the other, but I am not sure. Any references to papers/books are also welcome =)

Thanks in advance,

Pedro