Research And Implementation Of Surveillance Video Synopsis Generation Technologies

With the development of technology and safety awareness enhancement of the society, the surveillance cameras are widely used. With 24 hours working per day, these cameras captured millions of video. Due to the limitation of security guards resource and original weakness of human being, it is nearly impossible to find out all hidden danger when monitoring, therefore it is necessary to quickly browse and analyze the video after recording. Since video browsing is extremely time consuming, it caused most of the surveillance videos never watched and reviewed. Fast and completely browsing the long surveillance video become urgent requirement in the field of surveillance industry. However, the traditional video synopsis technique can not be adapted to the surveillance for it would loss activity. Based on the Chongqing's key scientific and technological project <> (No.CS7C2009AB2230), this paper proposed a systematic surveillance video synopsis technique framework. Surveillance video synopsis provides a short video representation, which preserving the activities extracted from the original video. The activity in the video in condensed into a shorter period video by simultaneously showing multiple activities. Key algorithms involved in the surveillance video synopsis are studied and implemented. The main tasks in the paper include:①Current mainstream video theme abstraction algorithms are described. The core algorithms in the synopsis framework are analyzed, which include motion detection and tracking, trajectories extraction and synopsis stitching.②The codebook model used to detection moving objects is simplified by assuming each codebook only has one codeword, and the data distribution direction is adjusted by using principle component analysis algorithm; With foreground information above, color feature is extracted for every moving target, and particle filtering algorithm is used to track multiple moving objects, furtherly obtain and save the trajectories data.③As the surveillance background scene is fixed, the tactics that shift the trajectories along the time axis is adopted. We described the energy of the trajectory itself and the interaction energy between trajectories. The modified simulated annealing algorithm is employed for the combinatorial optimization.④Gaussian distribution model is utilized to stitch the trajectory and the background image to generate one synopsis frame, which furtherly make up the video synopsis.⑤Designed an intelligent fast surveillance video synopsis generation and browsing system, and illustrated the detailed process of its technology framework.After the implementation of the entire procedure of surveillance video synopsis, the obtained synopsis showed it provided fast browsing while keeping the integrity of activity in the original video.