How to use robot_localization with ar_track_alvar?

edit

Both ekf_localization_node and ukf_localization_node assume that all measurements provide either the robot's body frame (base_link) velocity, its world frame (map or odom) pose, or its body frame linear acceleration. Of course, if your data is provided in another coordinate frame, that's fine, so long as a transform exists between the source and target frames. Your approach of using an intermediate node or modifying the existing node to provide the data you need is definitely the way to go.

Your second question is probably a better question for the Robotics Stack Exchange site. What you're asking is how to compute a covariance matrix of the robot's pose from a relative range and bearing measurement to a marker location, and I think the answer is sufficiently theoretical to warrant a question there. Off the top of my head, it feels like the error in the range and bearing from the robot to the marker is the same as the error in the range and bearing from the marker to the robot. If you can transform that range and bearing error to (X, Y, Z) error, then would it not be directly applicable to whatever pose measurement you compute? In any case, you'll have to be careful to compute the covariance in the world (map or odom) frame and not the robot's body (base_link) frame.

Just doing a brief look on Google, it seems like other people have applied the unscented transform to similar problems successfully.

In any case, the approach suggested by @Jim Rothrock will at least get you some result. If you went that route, you'd probably want to inflate the diagonal values a bit so as to over-estimate the error, though I realize you'll lose the correlation between the variables.