A 16x16 macro-block would be divided into smaller parts and each part have its own motion vector. If we were to define a representative motion-vector for the whole macroblock to be used in HVS analysis, what would it be?
The one that Damien suggests sounds good, but I see a problem with it:
When an MB is divided into smaller sub-MBs, it is very likely that the MVs corresponding to blocks are different. In this case, averaging (no matter weighted averaging or what) would probably result in an MV that does not correspond to any of the blocks. In other words, we'll probably face a totally new MV. This could be a problem especially when we have more than one moving objects in our MB.
For instance, imagine that we have two objects; the first one moving toward North-West and the second one moving toward North-East. In this case averaging will result in North. We all agree that North does not represent this MB's motion. This is the reason that neither AVC nor HEVC use "mean" for MVP (motion vector prediction). They use "median" instead.
I'd suggest using "median"; I don't know if there is any way for "weighted median" or not. But it seems useful if we do weighting as Damein suggested, but for median.
Some moving entities in the scene will be of greater interest than others. I think you should probably report the motion of the most important/salient object in a given "macro block". For instance: Much of the motion may be due to motion of the camera, e.g. panning, zooming, translating (of no interest); Some of the motion may be due to large regions undergoing uniform motion in the background, e.g. clouds, trees swaying in the breeze (of partial interest); Other motion may be due to threats in the foreground that are trying to destroy your camera (of great interest)!
So this approach requires some degree of scene understanding, involving tracking, classification, etc, which would involve a lot more computation and effort than the other suggestions proposed so far. It all depends on your application and goals I guess.