Studies On Moving Object Segmentation Method Based On Markov Random Field

In the field of image processing, the segmentation of moving object in video sequences is a hot research topic in recent years. Varies of segmentation methods can be used according to different situations of moving objects with their different background. Combined with color information, MRF(MRF: Markov Random Field) model, which focuses on the extraction of a few of objects in a stationary background, is a new simple and effective system of image segmentation and gives rise to a satisfying result.Key techniques are illustrated as follows:In the section of color clustering, the algorithm of region growing is used to get color information from a color image in RGB color space, and a follow-up algorithm is adopted to eliminate lots of tiny regions. The tree data structure is also employed to enhance the real-time ability.The First Scheme: An Improved MRF ModelThe traditional MRF model regards observations as prior knowledge, with which the optimized labels are obtained. The improved MRF model adopts the segmentation method of traditional MRF model: Firstly, two observations are derived from three successive images with the method of change detection. Secondly, the threshold chosen to detect the initial labels is obtained form the observations with the means method. Finally, the maximum posteriori estimator, which is determined oy using the iterated conditional mode algorithm, is employed to get optimized labels. The enhanced performances of MRF model contribute to the weakening of the noise and to the reduction of the covered-uncovered background and to the recovery of the uniform moving regions and to the convergence on the more accurate object contour.The Second Scheme: A New MRF ModelThe new MRF model presents a new method of image segmentation, in which the image segmented with the color clustering algorithm is regarded as prior knowledge to obtain the optimized label, and the corresponding Gibbs energy function is redefined. Firstly, two observations and two initial labels are derived from the three successive images with the same method in the first scheme. Secondly, the AND-label is obtained with the AND-operation on the two initial labels. Finally, the maximum posteriori estimator, which is determined by using the iterated conditional mode algorithm, is employed to get optimized labels. The new MRF model contributes to the weakening of the noise and to the elimination of the covered-uncovered background and to the recovery of the uniform moving regions and to the convergence on the more accurate object contour.The simulation results show that with the improved MRF model, the system is good at segmenting moving object with comparatively large inter-frame vector in a stationary simple background but it is bad in segmenting moving object in a stationary complex background.The simulation results also show that with the new MRF model, the system obtains more accurate object contour in a stationary simple background than the system with the improved MRF model. And what's more, the model is successful in segmenting moving object with very large inter-frame vector in a stationary complex background.