Moving Video Object Segmentation Based On Spatio-temporal Information

With the rapid development of computer technology and its allied subjects, digital video plays a more and more important role in our lives. The new coding standard MPEG-4 adopts content-based coding method, and it makes many scholars and experts to study the problem of video object segmentation in depth. As one of the basic problems of computer vision, video object segmentation has been used widely, such as video surveillance, video retrieval, moving detection, pattern recognition and control, etc. Video segmentation based on motion object becomes to be a hot research in the world.In this thesis, we make an in-depth study of Gaussian mixture model, and then an improved algorithm of moving detection is proposed, in order to solve the problem about illumination changes and the stability of the model. But the method based on the framework of time domain segmentation has its limitations. It can not accurately get the edges of moving objects. So we choose a method based on the framework of spatiotemporal segmentation. In aspect of spatial segmentation we use the EM algorithm and k-means algorithm to divide the similary pixels into their respective regions. Then we use multi-frame subtraction algorithm and canny edge detection in order to get correct object regions. Finally by combining the time domain location and the reuslt of spatial segmentation, we can get the video object.The method is based on spatio-temporal, so it contains more information about the object structure, and can get a more accurate result. Meanwhile, with the development of EM algorithm and k-means algorithm, we could reduce the time we need. Experiments show that the method can achieve good results.