Moving Object Detection Method Based On Markov Random Field

Moving object detection is the core technology for computer vision with the dynamic visual information perception. It is the basis of understanding moving object behavior and accomplishing higher level tasks. Moving object detection has been widely used in many areas. In this thesis, we study the moving object detection method based on MRF model. First we introduce a moving object detection method based on 3D spatio-temporal MRF and solved by belief propagation algorithm. Improvements we make on the compatibility function help to lower the influence of the thresholding over the initial label field acquisition and make detection results more stable. Then we propose an algorithm which combines the MRF model with Mode algorithm, considering the spatial relationship between pixels to improve the background updating process. We propose an optimization method on spatial domain according to the pair-wise relationship between pixels and the asynchronous accelerated message updating process. It can be used to optimize the detecting result from frame differencing, reduce the noise and complete the object hole. It has similar optimization results compared with spatial MRF model but with significantly low computation.