I'm not familiar with RTIMULib, but I can say that your assumption that you don't need to incorporate the sensor attitude into your motion integration is wrong. Also, Euler angles (roll, pitch, yaw) are generally not an adequate choice of attitude representation for a system that is expected to rotate freely about all axes, because you can expect to hit the singularities.
Essentially, what you need to do is called strapdown navigation (Let me google that for you).
There are numerous approaches for the initialization process. Most of them will begin with an alignment phase where a stationary condition is assumed. With that assumption and the accelerometer data, you can determine the direction of up/down by averaging a number of measurements. Then, combined with some averaged magnetometer data, you can determine the alignment of the IMU with respect to a more-or-less Earth-fixed frame (should be close enough for a ball being kicked around).
Once you're happy with the initialization, you can kick the ball and perform the strapdown integration to recover the ball's trajectory.
Beyond that, it depends on the specifics of your application. Are you just interested in estimating the current states of the ball at any given time, or are you going to kick, reset, and kick again (in which case you can make further assumptions and use more sophisticated calculations to incorporate them)? There is also sensor fusion and various filtering techniques which can help to reduce the error of your solutions.
Finally, don't forget that something as unpredictable as kicking a ball may require some attention to the measurement scale settings.
Hope this helps get you started.