Well, there are temporal samples, and spatial samples, and as far as i know, temporal samples are good for creating some motion blur, because they check previous frames too and somewhat overlaps with them.
Spatial samples just use your current frame, not previous ones. And if your use both, keep in mind, that the total sum of required/renderes samples per frame is the result of the multiplication of your number of spatial samples with your number of temporal samples (and if you use tiles from the high res shot, then multiply them too…).
F.E. 1 spatial sample and 1 temporal = 1 in total. 1 spatial and 32 temporal = 32 in total for one frame. 2 spatial and 32 temporal = 64 in total for one frame (and just for fun: with 4 tiles → 64 *4 tiles = 256 samples per frame). So balance things out between those samples.
So what happens, if you let the temporal samples at 1, and raise the spatial samples to 32 (isn´t 10 enough?)? Should reduce the blur, unless it´s created by your camera itself and is more depth of field than motion blur.
Also, there is that infamous “first frame bug”, which always makes your first frame looking completely bonkers, no matter what you might set as warm up times. There is a solution for that:
Also the advice, render just one camera cut in one go. If your sequence contains several camera cuts, then maybe create a buffer zone and discard those crazy frames later by hand. Bothersome if you have a dozen cuts, but thanks to that bug, it´s probably safer to make several small render sessions, and combine them later 
I have the suspicion, that they come from motion blur/temporal samples, since that checks previous frames, and if there is no previous frame, then it creates one with completely false camera positions and overlaps that with your correct first frame camera position. Could be happen for the last frame too, haven´t tested that.