| 摘要 |
An image analysis system generates a motion vector field from first and second images in a sequence of steps. First, random, initial motion vector values are assigned to each pixel in the image frame. Next, an objective function is defined for each motion vector and its corresponding motion vector in the other frame. The differences between the motion vectors of the target pixel value in the current frame and the other frame are minimized by minimizing the objective function values, first based on all neighboring pixel values. Next, the motion vector differences are further minimized by reducing the objective function values by rejecting motion vectors which correspond to the neighboring pixel values for which the value of the difference between the respective objective functions is outside of a threshold range. In one embodiment of the invention, pixel values in the first image frame are compared to neighboring pixel values in a second image frame which occurs after the first image and to other neighboring pixel values in a third image frame which occurs before the first image frame. This embodiment substantially eliminates motion vector errors due to both motion discontinuities and occlusion of pixels in the image. <IMAGE> |
