I put this to a friend of mine at work today, and he explained this is a problem with cameras that use cheap CMOS sensors. The problem is known as a rolling shudder.

It is indeed that, rather than a simultaneous reading of the electrical values on the sensor chip, on the cheap chips, it is read serially from bottom to top, so if there’s motion, the upper parts of the picture are from later in time than the bottom.

Drive past a vertical pole or tree and snap a picture as you pass and it will appear to be at an angle in the picture.