Research For Historical Feature Model Based Background Subtraction And Experimental Analysis

The inspection of foreground object is the first step of application of the automatic vision surveillance. In most cases, we screen the actual scene with video, then depart the locomotory object (foreground object) which we want to analyze from the relative static object (background object) in the sequential videos. This is the technology called "The background subtraction". This technology needs very sensitivity and preciseness. It can be used in some application of intellective vision, for example, the traffic surveillance, the garage stakeout and so on. At the same time, for some research of special vision analyze and vision detection, the background subtraction can also be used as a good accessorial technology.In this paper, we have integrated the thought of some classical background subtraction technology includes the single Gaussians model, the mixture Gaussians models, the non-parametric model, the codebook model and so on; and present an innovative pixel-based way for the background subtraction. We track record every pixel's current grey intensity real-time. But our method is distinct with the non-parametric about the tracking record all the history grey intensity of every pixel. We just classify all the history record of the pixel according as the different range and the different change rules of the pixel, and then build some classified models for every pixel correspondingly. The different model means the different classified history record, and implies the different sense of actual vision. All built models includes the two kinds of the model: the background model and the non-background model. And every model can update self-adaptively and real-time. We endow every model with a priority which can decide whether a model is the kind of background model or the kind of non-background model and can decide the conversion of the two kinds of model. Our algorithm says that any pixel is determined as the background object pixel if it can find a model of the kind of background model in its all built models, otherwise is the foreground object pixel. We also present an method of updating the model priority which is different with the way of the mixture Gaussians.We have done experiments of many different kinds of the scenes, and compared with the mixture Gaussians models and the non-parametric model. From all the results of these experiments, we have found that the our method can get the more clear-cut foreground object than the other two methods, and in these complicated and diverse scenes, our method can get the best result. At last, we use our method in the research of tracking of the human face and the motion analysis to prospect the future research point.