Research Of Moving Object Detection Based On Sparse Representation

As a pretreatment means, moving object detection plays a very important role in machinevision applications such as smart space, video surveillance, etc. Moving objects in objectdetection is called the prospect, which we interested in, however the other is calledbackground, which is to be ignored. The foreground region extracted in a video stream isusually prepared for the next step such as object identification, tracking, behavior analysis,which is the purpose of moving object detection. In this paper, the theory is applied in movingobject detection based on the in-depth study of sparse representation. Sparse coefficients forbackground can be obtained through the K-SVD learning dictionary or solving the RPCAmodel, meanwhile background update and foreground objects segmentation could be realizedbecause of the character of video sequences of correlation of time and space, in order torealize the object detection. The main work and research results are as follows:(1) Background and significance of the study and the application fields of sparserepresentation have been introduced in detail in this paper, which is aiming at establishing thetheoretical foundation for the follow-up paper algorithm.(2) In-depth study of the sparse representation theory, several common sparsedecomposition algorithms are described in detail. Meanwhile, we introduce SVD algorithmand K-means algorithm which are involved in K-SVD dictionary learning methods. Anddetailed descriptions have been involved in the principal component analysis (PCA) and therobustness of principal component analysis (RPCA) and their applications.(3) An improved algorithm of moving object detection based on K-SVD dictionarylearning is proposed. Firstly, the initial background image is got by multi-frame averagingalgorithm from the training samples, and then the initial background sparse representationmodel is built upon it by BP algorithm. Secondly, combining with the current adjacent fiveframes, the dictionary is updated adaptively by K-SVD method in order to make thebackground model approximate adjacent frames background’s observation values optimally.Finally, the foreground moving object is obtained by subtracting the background model fromthe current image.(4) A block-sparse RPCA algorithm of object detection based on PCP is proposed. Firstly,regarding the observed image sequence as being made up of the sum of a low-rankbackground matrix and a sparse outlier matrix, we solve the decomposition by the RPCAmethod via PCP. Then combining the consistent optical flow of motion saliency with thespatial coherence on these regions, the rough foreground regions are obtained. Finallyblock-sparse RPCA algorithm through PCP is used to estimate foreground areas dynamically, reconstructing the foreground objects. Extensive experiments demonstrate that this methodcan exclude the motion and change interference of background significantly and improve thedetection rate of small objects simultaneously.