Spatiotemporal multiresolution associated to MRF modelling for motion detection

We are concerned with motion detection in image sequences acquired with a static camera. The monoresolution motion detection algorithm based on spatiotemporal Markov random field (MRF) modelling was proposed by Luthon and Caplier (see 4th. Eurographics Animation and Simulation Workshop, Barcelona, Spain, September 1993). Our aim is the use of a classic multiresolution approach, that is to apply the motion detection algorithm on a low-pass pyramid related to a coarse-to-fine strategy. But we propose to implement this framework both in space and time and to build a spatiotemporal pyramid related to the spatiotemporal MRF model. The same low-pass filter is applied in space and time yielding the same model energy at each level. However, the lower the resolution, the weaker the interactions between the pixels. Since the good behaviour of the algorithm is based on a balance between the influence of each term of the energy, reducing the spatiotemporal interactions between pixels is equivalent to increasing the adequation energy along the pyramid. A description of the motion detection algorithm in the monoresolution framework is given. Then the building of the spatiotemporal pyramid and the application of the MRF model on such a pyramid are described. Finally, some promising results in the case of uniform moving objects and sub-pixel motion and some aspects about the computational complexity of the algorithm are presented