Yaw motions show up as roll motions

[bhomberg]

In using this driver for our UM7, we noticed that the data that's output isn't as expected. I collected a bagfile and plotted the data (using tf.transformations to switch from euler to quaternion). The robot is only moving in 2D; the IMU is mounted parallel to the floor (as in the Hardware Installation section at http://wiki.ros.org/um7). All of the non-noise motions show up as roll, though with the same magnitudes as expected from yaw, as calculated from our mapping framework primarily using lidar data:

Has anyone else seen this issue? We're using the standard tf_ned_to_enu parameter. Our version of the sensor is redshift labs' RSX-UM7 and is running firmware version U72B.

I'm going to make a patch for our system that handles the quaternions correctly for our sensor -- I am happy to upstream it if that would be helpful. I will likely add an option that outputs the IMU data in the sensor frame; we'll handle that by having the urdf know that the IMU frame is flipped upside down.