Illumination invariant motion estimation for video segmentation

In this paper, we develop a motion estimation method which is derived from an illumination invariant to detect video shot changes. Compared to traditional optical flow methods, our method can discount illumination changes. A condition number will be also calculated when we estimate the motion vector at a certain point of the image. Two measures are defined: one is the mean square deviation of estimated motion vector for all the pixels in the same image, the other one is the statistics of all the condition numbers in the same image. Either of these two measures corresponds significantly to some types of shot changes. The experimental results and performance evaluation are presented to demonstrate the efficiency of our method.