The Research On Foreground Processing And Content Exhibition Of Surveillance Videos

There are so many surveillance cameras spreading all over the world nowadays. Wherever you go, a camera shot which locates in a very corner may catch you and millions of surveillance cameras record massive data every day. Therefore, how to deal with the video data has become a big challenge to the users. It is also a new research area for the academia and industry. This thesis discusses the problems involved in video content analysis and processing. Three algorithms are proposed.First, surveillance video content analysis is introduced. An algorithm is proposed to separate foreground objects from background. Based on Gaussian mixture model, an automatic method is used to do the segmentation work. Several changes are introduced to improve the performance of GMM, which not only extract all the foreground objects, but reduce the storage space of videos as well.Second, the similarity of foreground objects is estimated. Through extracting kinds of image features and matching them, the score of similarity between videos can be got. Since the same object may appear in different cameras at different time, it is very useful to find out the relationship between different videos. This procedure also builds foundation for the next step which is video collage.Finally, a dynamic video collage approach that summarizes multiple dynamic activities in parallel on the display canvas is presented, which is also the key of this thesis. Most of existing image/video collage methods produce static key frame based collages, which only contain limited information of videos and may influence users’ understanding of visual datasets greatly. The long processing time due to large amount of optimization and software coding only is unacceptable. This thesis proposes to utilize activities cuboids to conveniently reorganize dynamic objects into collage elements. Spatial-temporal optimization is then carried out to optimize the positions of the activity cuboids in the 3D collage space. This thesis facilitates the efficient dynamic collage via event similarity and moving relationship optimization on CUDATM platform allowing multiple video inputs. In multiple experiments and user study, the method consistently demonstrates the efficiency and effectiveness of dynamic collage.The main contributions of this research work are as follows:· Add new mathematic models to the segmentation algorithm which can weaken the bad effects made by illumination change and camera shaking. It can produce more accurate segmentation result.· Through similarity estimation of foreground objects, the relationship between videos from different cameras can be found.· A dynamic video collage algorithm is introduced. The video collage with kernel recording stream GPU processing enables condensed dynamic summarization for easy browsing of long surveillance videos while saving memory size for storing the huge datasets.