Research On Moving Target Detection Algorithm Based On Non-negative Matrix Factorization

As an important application of computer vision technology, intelligent video surveillance system has been one of the active subjects in computer vision field. Without human intervention, intelligent video surveillance system analyses and processes the video sequences in real-time and automatically, attempting to detect the moving target objects in video sequences, thentracking, recognizingand analyzing the subsequent behavior, so as to understand and explain the video content, and make real-time early warning and proactive behavior guide. As the basis of video monitoring system, the accuracy of moving object detectiondetermines the reliability and accuracy of the subsequent steps.This thesis mainly studies the algorithms of detecting the moving objects in the intelligent video surveillance system. The non-negative matrix factorization(NMF) algorithm is introduced into the moving target detection algorithm for modeling the background of a video sequences, and the background subtraction method is used to obtain the approximate foreground image by comparing the differences of the current video frame and the background model. Finally, the complete information of the moving targets is obtained by the post-processing, such as thresholding, morphological processing, etc.Firstly, the origin, the basic mathematical model, cost function, multiplicative iterative rules, basic steps and convergence of the basic NMF algorithm are introduced in this thesis. Meanwhile, we find that the algorithm with a batch process is not suitable for the real-time application because of its high computational complexity and space complexity. So in order to solve this problem, this thesis proposes a sliding window based non-negative matrix factorization moving target detection algorithm, which control the size of the decomposed matrix in NMF matrix decomposition model by adjusting the length of the sliding window. The proposed algorithm can reduce the computational complexityand space complexity, and to some extent, it can increase non-memory characteristicof the model.Secondly, aiming at the deficiency of the basic non-negative matrix used in the real-time applications, according to the incremental principal component analysis (PCA) algorithm, this thesis apply the incremental non-negative matrix factorization algorithm to solve the moving target detection problem. The origin, mathematical model, cost function, multiplicative iterative rules, algorithms and basic steps such as convergence of incremental non-negative matrix factorization algorithm are described in the thesis. Meanwhile, the analyzed results of the computational complexity and space complexity indicate that they can meet the real-time requirement of moving target detection applications. As the solution of the basic non-negative matrix factorization algorithm is not unique, this thesis presents a model based on incremental sparse negative matrix factorization of moving target object detection algorithm, which reduces the computational complexity and the space complexity of the algorithm by the incremental model, so that it can meet real-time requirements. Meanwhile, the algorithm can control the contributions of each sample model, so that it can enhance non-memory characteristic of the model and the ability of adaptive changing for the dynamic contents. The experiments show the good detection results in the complex environments with illumination and background changed.