Research On Surveillance Video Sparse Representation And Dictionary Learning Based On Compressed Sensing

In recent years,with the development of multimedia technology,surveillance video plays an important role in our lives.However,massive data that comes with the popularization of surveillance videos challenges the computer storage and processing capacity.The development of surveillance videos in the future will be affected by two key factors,which are compressing and reconstructing videos efficiently.Thus,research on compression,reconstruction,representation and analysis of surveillance video efficiently is an important and valuable work.It is characteristic that the background in surveillance video is static or changes very slowly,and combining with compressed sensing system,an efficient sparse representation framework with background subtraction is proposed in this thesis.There are problems that image patch needs to be transformed from matrix to column and the whole training set needs to be used at one time in classical dictionary learning.Thus,an adaptive dictionary learning algorithm called block recursive least squares(BRLS)is developed to train dictionary with using image block directly and successively.A novel framework that uses multiple dictionaries for ensemble dictionary learning is proposed to improve the performance of sparse representation in frames with large size.An optimization method that designing one sensing matrix within multi-dictionary system is proposed.In addition,the semi-tensor product is developed in multi-dictionary framework to reduce the storage and optimization time when designing sensing matrix.The main contributions of this thesis are summarized as follows:1.A novel framework with background subtraction is proposed for surveillance video sparse representation and reconstruction,and adaptive dictionary learning algorithm BRLS is developed to this framework.Thus,the image patch can be used directly without transformation between matrix and column so that the information among pixels will be reserved.Besides,training frames can be used one by one rather than inputting the whole set at one time with high computational burden.2.Specific to surveillance video frames with large size and rich contents,a multidictionary learning and ensemble sparse representation scheme is designed.Simultaneously,an optimization algorithm is proposed for one sensing matrix within multi-dictionary framework.3.For reducing the amount of data and optimization time of multi-dictionary system,semi-tensor product is developed to this system.Thus,the sensing matrix can be designed beyond the limitation of signal dimension.At last,block multi-dimensional sparse representation is developed for colorful surveillance video frames.4.All methods mentioned above are tested with real images and surveillance videos with Matlab software.Compared with the classical methods,the effectiveness and advantages of these new strategies are demonstrated.