How can a NN guess the velocity of an object from a video source?

In the case of a video stream, I’d like to detect the speed (an approximation) of an object, that is moving. What would be the best approach to take?

I am thinking of 3 methodologies to take, though I’ve not found a lot resources to read about.

Method 1: I could feed Frame images to my CNN and pinpoint the object in the frame. Then calculate the difference of the position boxes.

Frame 1: [  o     ] ---> NN --> 3
Frame 2: [      o ] ---> NN --> 7

7-3 = 4 -> so dX = +4 between 1 frame, so I can estimate the dT (if I know the sampling rate of the frames)

Method 2: Is there any way where the NN can take as a feed-input the context of the previous frame and make the calculation itself?

video lapse 1sec: [  oooo ] ---> NN --> 4m/sec

Method 3: What if I can control the shutter speed of my camera, could I calculate the velocity of an object by the motion blur?

Frame image: [   ---o ] ---> NN --> 4m/$shutterSpeed

Any relevant resource to read would help a lot.