Research On Video Object Segmentation

In many computer vision applications, a fundamental and critical task is toextract foreground motion objects from video sequences. This technology is widelyapplied in human-computer interaction,3D reconstruction based on video, intelligentvideo surveillance, video conference, Intelligent Transportation and so on.At present, there are a lot of video moving object detection methods about thefixed camera. But when the foreground and background color, brightness is near tendto cause miscarriage. The scene illumination change would also cause some mistake,and almost all of the methods would be affected b y the shadows. If the aboveproblems can’t effectively solved, it would affect the follow-up related application,such as moving object tracking, event detection, behavior analysis, discriminatoryanalysis and so on. Although there are some methods improved the precision of theextraction of motion objects, because the used models are complex the computation islarge. So these methods can’t meet the requirement of real-time video processing.Therefore, how to build a kind of robustness method to extract the motion objectseffectively is very important.In this paper, we combined the image pixel level and regional level features, weproposed a video object segmentation method based on digital matting framework.And a method based on characteristics of joint modeling. By used these methods wecan get more accurate results. At the same time, based on the object segmentationalgorithm proposed in this paper, some work on vehicle detection and vehicle typerecognition is done, and a highly efficient recognition algorithm is put forward. Thepaper’s main work are as follows:1In video object segmentation we introduced GrabCut technology to extract theforeground from natural scenes. The temporal and spatial information of pixels areused. 2GrabCut technology need a lot of calculation, in order to save storage spaceand reduce the computational cost, the SIFT key points detected method is utilized.3In order to improve the quality of foreground segmentation, we use variousimage features to build corresponding gaussian models and discriminate foreground.4In order to improve the accuracy of vehicle type recognition, we use somedifferent image features, considering the edge gradient information and the globalcontent information of moving vehicles images. We use support vector machine(SVM) to effectively learning and training. At last we use the evidence theoryalgorithm to fuse the forecasting results and obtained good experimental results.