I would recommend you start with doing just head motion correction. The effect of head movements on FMRI data can be huge, and so removing them (as much as possible) is important. In particular, since the head moves as a unit, head motion produces highly correlated signal changes across wide regions of the brain. In FMRI processing, it is common also to linearly regress the 6 estimated movement parameters out of each voxel time series to further reduce the impact of motion.

For your purpose, you might not want to do slice timing correction -- you'll have to experiment to see how big an effect it has on your results. The same applies to spatial smoothing -- which of course blurs the signals between the voxel time series.

If you are going to be combining data across groups of subjects, then you need to align (register) their brain datasets together as far as possible. If you are dealing with only 1 subject at a time, then you can avoid this step, at least at first.

If you are planning to use AFNI, and you have an FMRI dataset in the NIfTI format (e.g., from OpenFMRI), you can do the volume registration with program 3dvolreg. If you want to further project out the motion parameters, program 3dTproject can do that for you. Program 3dmaskdump can then dump out all the data time series to ASCII text (which will be HUGE) and then you can take it from there.