Research Of Moving Object Detection Algorithm Based On Kernel Density Estimation

Intelligent video surveillance is an important and common application in the daily life. Systems of intelligent video surveillance usually use fixed sensors to monitor an area with the goal to automatically understand what is happened in the site. Since moving object detection is the previous stage of further processing in intelligent video surveillance, it has an important value in both research and application fields. So that many researchers paid attention to it and proposed different solutions.For the purpose to realize moving object detection by modeling the real world more reasonable, strategies based on kernel density estimation of nonparametric statistics have been proposed to estimate the probability density function of background pixels’feature in recent years. The kernel density estimation estimates the probability density function of pixel’s feature directly from the sample data without any assumption about the underlying distribution. Handle the arbitrary shape of pixel’s probability density function. The main achievements we got are listed below:(1) We present a foreground feedback strategy to construct a background and foreground modeling using kernel density estimation. For the current image, firstly, the model uses the basic kernel density estimation to get a probability about how possible a pixel belongs to the background. Secondly, the model uses the foreground area in the previous detecting result to build kernel density estimation on each pixel in the current image, getting a probability about the pixel on what distant belonging to the foreground, this aids the object detection. Judge each pixel in the current image whether belongs to the background or not by combining the background and foreground probability together. This proposal use previous acknowledge sufficiently while keep a quite acceptable computational cost, raising the efficiency and reliability of the detecting algorithm.(2) An adaptive background updating method using foreground edge detection has been given. It raised the detection accuracy while successfully handled the illumination change and ghost shadow in the background. We count the time when a pixel appears as the edge of the foreground. If there is a pixel staying as the edge of the foreground in a time period, we think that there is scene changed in the background, then rebuilding the sample dataset. Otherwise, just update the sample dataset.(3) An exploration on shadow elimination algorithm based on kernel density estimation in the HSI color space is given. Different from directly using pixel’s vector values to detect the shadow area in the current algorithm, we tried to give another way to realize the shadow detection by combining several issues about the probability density function values got from kernel density estimation to detect the shadow area. It achieves certain effect.(4) Completely analyses the theory and application on kernel density estimation. Give a direction for how to optimize kernel density estimation. So that it can be used in engineering applications more conveniently. We give our proposal:a newly constructed piecewise linear optimized kernel function. This piecewise linear kernel function have the smooth characteristic of Gaussian kernel function but lower computational cost when it is used in1-dimentional space, suitable for engineering applications.