A B S T R A C T This paper addresses model based object oriented motion estimation from image sequences. A generic label field segments the scene into several continuously moving 2-D objects. An image model assuming segmentwise stationarity of the displaced frame difference (dfd) and of the estimated fields is proposed. The dfd is shown to obey a white generalized gaussian distribution better than the commonly assumed overall white gaussian distribution. A coupled weak smoothness constraint bounds the segments of the label field to smooth shape and the vector field to smoothness within each of those segments. The MAP-estimator with respect to the image model is derived. Its performance is demonstrated by experimental results.
I N T R O D U C T I O NMotion compensation is a key technique for efficient video compression. Block oriented motion compensation is easily implemented but has several well known drawbacks. E.g. blocking artifacts are introduced since the block segmentation does not correspond to the object shape in the scene.Object oriented motion description aims to adapt the segmentation to the scene such that each region uniquely corresponds to one continuously moving 2-D object. Hence several drawbacks of blockwise motion description are avoided with the object oriented approach. However, object oriented motion estimation states an ill posed problem. This paper focuses on the formulation of an appropriate joint estimation criterion for the vector field and its segmentation.An early parametric gradient estimation technique minimizing the MSE of motion compensated prediction was presented in [7] and extended in [2].Another class of estimation techniques is based upon regularization theory. In [6] the MSE-criterion is extended by an additive smoothness term ascertaining smooth motion fields. In order to account for motion discontinuities, the smoothness constraint is relaxed along image intensity gradients in [9]. In [8] a model based approach is proposed which incorporates a hidden line process accounting for motion discontinuities. The smoothness constraint for the motion field is suspended across line elements. An estimation criterion consisting of additive terms accounting for MSE, motion discontinuities as well as for uncovered regions is investigated in [3]. The individual terms as well as their relationship are more or less heuristically imposed. All abovementioned approaches include the mean square of the displaced frame difference (dfd) in their estimation criteria.This contribution follows a Bayesian approach based on a statistical image model. It considers dfd, motion discontinuities and uncovered regions. In the following section, a new model for the dfd is discussed. Then, the a priori model for the motion field is formulated. It employs Gibbs/Markov random fields for region modelling which were introduced in another context in [l] and have proven suitable for modelling of images. The resulting estimation criterion is derived in a straight forward manner as MAP-criterion according to the model...
