Research On Motion Segmentation Based On Over-relaxation Algorithm And MRF

Segmentation and track technology of motion objects are always important subjects in image processing and computer vision. Recently, with the fast development of multimedia technology and continuous improvement of computer performance all kinds of video images are widely applied. Segmentation technology of motion objects in video image sequences attracts more attention and is used increasingly.After reading articles in this domain this thesis summarize development and research status of motion object segmentation. Firstly, related concepts and basic principles of motion estimation are introduced in the thesis .Then Markov random field (MRF) algorithm is presented in detail. A motion segmentation method combining motion estimation with MRF model is put forward. Firstly, synchro estimation of multi-motion and over-relaxation algorithm are used in motion estimation to obtain initial regions of motions. The robust formula is combined with error model and affine motion is selected as motion model. Motion parameters of every motion are achieved by over-relaxation iteration. The least error motion of each pixel is taken as the respective motion to which it belongs. The collection of all pixels belonging to one motion is the initial value of corresponding region of the motion. According to the characters of MRF of few parameters of model, strong spatial constraint and easy combination with other methods, MRF model is used to smooth the result and reduce noise. The simulating implementation of the algorithm and results of experiments are given in the end of thesis. The experiments show the method in this thesis can get better estimation and segmentation of motion objects.