Samenvatting

This paper describes segmentation based on motion. A group of (connected) pixels that move together is assumed to belong to one object. One advantage of using otion is that an object that consists of different parts is detected as one object, whereas f for instance colour differences are used the same object may be detected as several smaller segments which may have to be put together by sophisticated algorithms. Another advantage of the method is that after the motion is estimated, it becomes possible to correct for the changes due to the motion and then average the corrected frames in order to improve the signal to noise ratio. In this paper, a brief description is given of the original motion estimation method described by [Odobez and Bouthemy, 1995]. This method is very well suited for the estimation of the camera motion. Experimental results obtained with this method from real infrared and colour videos are presented. The results are very accurate, even when noisy videos are used. The same method can be used to estimate the motion of separate objects as well, by using only regions that move differently than according to the cameramotion. This extended method is described in detail and the results when applied to synthetic and real sequences are discussed. After having found the motion parameters of an object, the exact location and the shape of that object as it appears in each video frame is known. Therefore, the extended method can be used for segmentation based on motion.