Here are two possible solutions to your problem:
1) Launch both the camera and tracker normally, and rosbag record /tf (as you've done), which will have all of the messages published to tf. You can then filter out unwanted messages using rosbag filter.
rosrun openni_tracker openni_tracker
rosbag record tf
rosbag filter [recorded_bag] [filtered_bag] "m.transforms[0].header.frame_id == '/openni_depth_frame'"
This will filter only messages with the '/openni_depth_frame' as the frame id (leaving only the skeleton frames, which are children of /openni_depth_frame).
2) Remap the tf output of the openni_tracker to a different topic, and then rosbag record that topic, which will contain only the transforms from the tracker node.
rosrun openni_tracker openni_tracker tf:=tracker_tf
rosbag record tracker_tf
In this case, the skeleton frames will not be published to /tf, and so will not be visualized in rviz as part of tf, if that is a concern.
