A blurred image is equivalent to an image that is convolved with for instance a Gaussian kernel. This means that low frequencies have become favoured over high frequencies. A relatively blurred image would have suppressed high frequencies with respect to an image that you view as not blurred. Hence what you could do is study the typical high frequency content in normal images (e.g. Fourier transform based) and see if you can detect the blurred images that way (since they would lack some of the high frequencies). So take the frame by frame Fourier transform and study the alteration of high frequencies as a function of time.