Dynamic Scenes Video Segmentation Algorithm

Video segmentation is the technical foundation of many object-oriented video applications, such as object-oriented video encoding/decoding, video edit and index, and computer vision. The accuracy of segmentation directly affects the success of the next task, so video segmentation is one of the most difficult parts. Especially after MPEG-4, an object-oriented video compression standard, had been proposed, video segmentation becomes a exigent problem.In 1990s, many scholars began to do research on video segmentation, and proposed many methods. Since the complexity and particularity of the video image, the existing theories and algorithms for video segmentation still have some drawbacks and can not segment video object satisfactorily in some case. It is hard to propose a general method of video object segmentation applied to all cases. So, it has been the focus of current research work to find new methods for video object segmentation with specific application requirements or to make improvements to existing methods.Based on the information used in the segmentation process, existing extraction methods can be classed into three categories: temporal segmentation, spatial segmentation, spatial-temporal segmentation. For better segmentation effect, current mainstream video object extraction algorithms combine temporal field's continuity and spatial field's similarity to get semantic video object. To enlarge the information in the video streams, the camera needs to do movements, i.e. pan, tilt, rotation and translation. Camera's movements lead to dynamic scene video sequence. Extraction of video object from dynamic scene has many practical applications in video tracking, video summary and digital surveillance.Due to the movements of background, direct extraction becomes difficult. Newly proposed methods of video segmentation in dynamic scene do global motion estimation first to compensate camera motion, then use traditional still background method to extract video object. Precise back- ground compensation can bring large computation. In the case of considering practical applications, we propose an algorithm to extract moving object from dynamic scene based on spatio-temporal information. In temporal field, a coarse motion analysis method is used: First we find out motion vector, and then global motion estimation with parameter-model is used to find the foreground area. Then morphological dilation methods are selected to obtain final motion template, make sure the template contains the whole foreground object. And then spatial segmentation using graph pyramid is applied to image region in the current motion template. Through the method we can get better results with less computation.The proposed method avoids precise background compensation and is very computationally efficient, while the extracted semantic object is of high precision. The experimental results show that both rigid and non-rigid moving objects in dynamic scene are well extracted by this algorithm.