Study On Non-Parametric Background Modeling Method

Background modeling is the basic part of object detection and tracking, which is widely used in intelligent surveillance, robot vision, and video conference. Parameterized Gaussian mixture model is one of the best models for modeling a background scene with gradual changes and repetitive motions. However, it fails when the scene changes suddenly and false or missed detection caused by shadow and noise. In addition, this method needs to know the prior probability and conditional probability of the background. However, the probability function form is unknown in some complex application scenes. While non-parametric background modeling method does not need to presuppose the form of the model density distribution. It can directly estimate the unknown density function by the video data. This thesis investigates the methods for non-parametric background modeling, the main work are as follows:1. Some popular adaptive dynamic background modeling methods are deeply studied and the advantages and disadvantages of each method are summarized.2. The Gaussian mixture background modeling method is deeply investigated. A new Gaussian mixture background modeling algorithm which can reduce the shadow is proposed. The shadow detection which is based on the HSV color characteristics is embedded in the algorithm. It can reduce the noise and shadow and improve the robustness of the moving object detection.3. The non-parametric kernel density estimation background modeling method is also investigated. The model is set up and updated based on the Gaussian kernel functions and window size. It can adapt to the changes of environment, weather and illumination and can correctly detect the indoor and outdoor moving targets. Experimental results show that the non-parametric method has less noise, clearer targets and better connectivity than the Gaussian mixture background modeling.