Research On Moving Object Detection Algorithms Based On Mixture Gaussian Model

With the development of computer science and communication technology, as well as people's safety consciousness enhanced increasingly, more and more cameras are installed in the parking lot, the transport hub, the market and the public place. Since the operation of these cameras, which will produce a wealth of original information, they might be monitored by hand or be stored. And the information is a very tedious work by the manual processing, because it contains many things which are not our concern. It requires researchers to develop algorithms for the machine to help them complete the target detection and extraction.Currently, moving object detection algorithm is mainly divided into three categories:optical flow method, frame difference and background subtraction. And background subtraction is a mainstream method. The basic idea as follows:first, it need build a model which is close to the background, then making a new frame minus it. Finally, according to pre-set threshold, the pixels which do not meet the background model will be found. These can be as the prospect. Therefore, background modeling is crucial importance in this method. And Mixture Gaussian Models is one of popular modeling in the extraction of target, which has drawn wide attention. However, it has a large calculated amount. And the moving object will be turned into background even disappeared if it stay longer at a point.To solve these problems, the paper has done some study and the main work is as follows:(1) In the classic algorithm of Mixture Gaussian Model, because background modeling is based on the pixels, and it has three models in the pixels, in order to reduce the impact which changes external factors. The image will be partitioned into granularity coarser than pixel by a suitable equivalence relation in this thesis, which uses the advantages of quotient space hierarchical theory in the face of solving problem. The attribute [f] of quotient space([X],[f],[T]) can be acquired. Then model will be built for new granularity. It can enhance the speed of moving object detection.(2) The moving object when it stays one point too a long time will be turned into background or will be disappearance in the Mixture Gaussian Model. So it will be discontinuity in the scene, and it may affect the object tracking and positioning. The thesis tags the object detected by using a rectangle, and then the time which object stays in the scene are determined. If this occurs, we merely update models of the background region outside the moving object. Otherwise update the whole frame's model. So that it can maintain the continuity of movement, facilitate object classification and identification, behavior understanding, motion analysis and other advanced processing.