A Study On Video Synopsis Algorithm In A Complex Background

Video synopsis is one important research branch of intelligent video analysis.In the video synopsis algorithm,a relatively small number of key frames can be extracted by analyzing the structure and content of video sequence.As a result,the redundancy can be reduced significantly.Meanwhile,most content can be effectively retained for the new video sequence.In a complex background,e.g.a low-light condition or leaf disturbance,how to correctly detect the foreground target and improve the video concentration rate is a difficult problem.Aiming at these issues,we have carried out the following two works.(1)An effective video synopsis method is proposed for the surveillance video under a low-light condition.First,a low-dark background is transformed into a fog-like scene by applying an inversion operator on each pixel of the video frame.As a result,the low-lighting image enhancement problem is changed into a dehazing problem.Then,an atmospheric scattering model is constructed,and an optimized transmission is computed by means of the dark channel theory and the guided filter.In order to improve the computation efficiency,a selection strategy based on the hash value is devised to determine if it is necessary to recalculate the transmission.Furthermore,an improved vibe algorithm is used to model and extract foreground objects.A strategy based on the difference between frames is proposed to eliminate the ghost.According to the ratio between the foreground object and the whole image,we judge whether there is a moving object.Finally,a condensed video is constituted of the frames with a moving object.(2)An effective video synopsis method is proposed for the surveillance video when there are shaking leaves in the background.First,the moving object is separated by constructing a Gaussian mixture model for the input image.And the noise is removed by means of some morphological operators,such as eroding and dilating.In terms of the proportion of the foreground object to the whole image,a preliminary judgment is performed on whether retain the current frame or not.Moreover,we divide the background frame and the current frame image into small patches and calculate the corresponding histogram differences.Then,the leaf disturbance or the foreground target can be distinguished by the histogram differences.According to the ratio between the foreground object and the whole image,we judge whether there is a moving object.Finally,a condensed video is constituted of the frames with a moving object.Two video synopsis methods presented in this thesis have been tested on some publically available datasets and the videos sampled by ourselves.Compared to several existing algorithms,the experimental results demonstrated that the proposed methods are more robust to the low-light condition and the leaf disturbance.