Study On Video Moving Object Segmentation Based On Spatio-Temporal Information

Video moving object segmentation, which is one of the key supporting techniques for the MPEG-4 content-based coding scheme, is widely used in the fields such as video surveillance, machine vision, intelligent transportation system, weapon guidance, etc. It's the hot issue of digital image processing technology and multimedia technology, and has great significance both in theoretical research and practical application.The main content of this thesis is the video moving object segmentation based on spatio-temporal information. Segmenting the moving object from the video image is a research problem that detecting the non-cooperative targets from multi-dimensional signals with non-stationary characteristics. The random property of this problem in time-space domain is analyzed. Following the main line of temporal motion information, the video moving object segmentation technique, in combination with the spatial information, is studied in detail. The main results are as follows:1. The Basic concepts of video moving object segmentation are introduced, the general model of moving object segmentation system is discussed, and the common classification features and classification rules are analyzed. Then the basic model of video moving object segmentation based on spatio-temporal information is constructed, and the advantages of spato-temporal method are studies.2. The stochastic contour model of moving object is studied, the contour detection theorem is proposed and the hierarchical contour detection system is established. The motion-changed region of the difference image is detected by binary clustering according to the global adaptive threshold, the high SNR section of contour is obtained by the per-pixel"AND"operation between the motion-changed regions of two successive difference images, the low SNR section is detected in the time-space domain by the occlusive criterion and Markov property. The moving object is segmented according to the contour of object.3. A novel spatial segmentation and video moving object detection based on fuzzy clustering is proposed. The local variance is chosen as the fuzzy characteristics according to the different properties of motion-changed region and relative noise region in the difference image. The fuzzy clustering criterion is established and the fuzzy partition is obtained by the Powell algorithm. The defuzzification is carried out, the motion-changed region and the moving objects are detected.4. An improved level set method is proposed. Considering the gray level distribution of pixels in image, a new energy function based on the error rate is achieved, and the corresponding evolution equation is obtained. Unlike the tratidional level set method, the improved level set method with the characteristic of global optimization can be applied to the images that the distribution of pixels is uneven, and it also can deal with the images with blurry boundaries or strong noise.5. An improved watershed in accord with the human visual characteristics is proposed. The noise is suppressed by morphological reconstruction, the gray level non-linear transformation for the morphogical gradient image based on Weber perception principle is carried out, and the marker-based watershed is fulfilled. The improved watershed can restrain over-segmentation effectively.6. A moving object detection based on Gaussian mixture model is proposed. First the difference image is modeled as the mixture of Gaussian distributions, the expectation maximization algorithm is carried out, and the temporal segmentation mask of moving object is obtained. An improved method for model selection and parameter initialization is proposed, and the model size for Gaussian mixture is quickly chosen. To integrate with the improved watershed method, the moving objects are segmented.7. A novel moving object detection algorithm based on Markov random field is propsed. To cosider the spatial distribution of pixels, the marker field of difference image is model as Markov random field. The MRF is iterative solved by mean field algorithm. The new model selection and parameter initialization is used,and the temporal segmentation is obtained. The spatial segmentation is achieved by the improved watershed algorithm. The temporal and spatial informations are fused and the video moving objects are obtained.