Moving Target Detection Based On Dictionary Learning

Intelligent video surveillance is an important branch of machine vision. It has beenwidely used in airports, banks, shopping malls and other places and it is the developmenttrend of the future security monitoring technology. Moving target detection is one of the keytechnologies of the intelligent video monitoring technology, at the same time it is the premiseand basis of target tracking, target recognition and the analysis of target behavior. On the basisof in-depth study of the video sequence’s sparse representation and dictionary learning theory,we apply them to the detection of moving target in this paper. Firstly, background sparserepresentation model is created through the dictionary learning, then conduct dictionaryupdating and target segmentation by exploiting the spatial and temporal correlation betweenbackground and foreground, and as a result the moving target is detected accurately. The mainwork of this paper is as follows:(1) The thesis describes the research background, research significance and study statusat home and abroad of the intelligent video security surveillance through intensively reading alarge quantity of domestic and foreign literature and writings, Simultaneously we make arelatively comprehensive review of classical and mainstream algorithms of the moving targetdetection, which lay a solid theoretical foundation on the research of moving target detectionalgorithm in this paper.(2) The thesis introduces the dictionary learning and sparse representation theory anddescribes the matching pursuit algorithm, orthogonal matching pursuit algorithm, basispursuit based algorithm and Focuss algorithm in detail and describes how they apply to thedictionary learning in the image and video sequence processing.(3) We put forward an improved target detection algorithm based on Dictionary Learning,which builds background sparse representation model via the dictionary learning. Inspired bythe K-SVD algorithm, dictionary atoms and sparse coefficients are updated at the same timeunder this proposed dictionary learning method, but the calculation process discards singularvalue decomposition, which can reduce the amount of computation. Meanwhile the dictionaryis divided into K classes, respectively, corresponding to the K-class characteristics of thebackground. The updating of the dictionary makes the background model be more adaptive toenvironment changes, after background subtraction of the current image and the updatedbackground, employ segmentation algorithm proposed in this paper for post-processing, thenthe moving target is extracted.(4) We propose an algorithm of target detection based on principle component pursuitand dictionary learning. The image is made up of three components in this algorithm, thestatic background, dynamic foreground and the noise obeyed Gaussian distribution. Therefore, pre-processing in video image is adopted in this algorithm at first, then employ principlecomponent pursuit method to generate the initial sparse dictionary. After that create the sparsebackground model through dictionary learning and encode sparsely for the video stream to bedetected. Finally background subtraction method and segmentation algorithm are used tocomplete the detection of moving target. It shows that this algorithm can detect the movementof vehicles and pedestrians accurately by means of highway surveillance video and car parksurveillance video experiments.