Research On Some Algorithm Of Video Moving Target Detection

With the rapid development of information technology, and the popularity of safetyconsciousness, the social demand for intelligent video surveillance is more and more high. As akey step, moving object detection with a long history of research and continuously changedresearch methods has become a frontier topic in many disciplines. On the basis of meeting thereal-time performance, robustness and accuracy, this paper embarks from different researchmethods to realize the moving object detection in different scenarios. According to the series ofproblems, this paper expounds from the following three aspects in turn:1、In order to remove the ghost, this paper put forward a improved method based on visualbackground extractor algorithm(Vibe). Firstly, analysis of the superiority and inferiority ofalgorithm, especially considering the detection results in the initial period of ghostingphenomenon, extraction with three frame difference method for initial background. Then usestochastic neighborhood sampling to establish multiple samples of background model. Secondly,distinguish the foreground and background by using the Euclidean distance of RGB color space.Finally, use “AND operations” with results and the current image to realize the remove of ghost.2、Achieve the shadow removal by texture information. Firstly, compare the characteristicbetween Local Binary Pattern and Scale Invariant Local Ternary Pattern, and then do re-texturingfor the current image by using local scale invariance of SILTP. Lastly, complete the thresholdsegmentation, which realize the detection and removal of heavy shadow.3、This paper put forward a moving object detection algorithm, which based on low-rankmodel and alternating direction method. Firstly, analysis the background modeling of RobustPrincipal Component Analysis (RPCA), and all kinds of classic methods of solving RPCAmodels has been analyzed. Principal Component Pursuit (PCP) is emphatically discussed, and therelated simulation experiments for PCP have been done. Secondly, the model for classic PRCAhas been improved by decomposing the actual data into clear data (of the background model),singular value data (related to foreground model and moving target), and noise data (ofbackground noise model). Add total variation constraints in the low rank background model toensure its relative stability. And adopt Markov Random Foreground Entry Point skills to thesparse model for the stochastic modeling to get an improved sparse low-rank model. Thenalternating direction is used for solving the model within the Augment Lagrange multipliermethod framework. Finally, the simulation results demonstrate the effectiveness of the method.