Researches On Background Modeling Method Under Complex Scene

In recent years, background modeling and foreground detection has been widely used in the field of computer vision, such as video surveillance, target tracking, human behavior recognition, license plate detection, etc. At the same time, the application environment of background modeling and foreground detection has also changed greatly from the static indoor scene to the outdoor scene which contains strong disturbance. Due to the complex scene contains dynamic factors, such as grass shaken by the wind, camera shake, illumination change, moving foreground, how to raise the speed of background modeling and reduce the memory consumption under the premise of high precision in foreground detection has always been the challenge of the method. From the perspective of processing units, background modeling and foreground detection can be divided into region-based and pixel-based method. The region-based method has high background modeling speed, low memory cost, strong ability to resist noise and strong jagged outline in the foreground detection result, is suitable for the computer vision applications with strict requirements on memory cost and speed such as video surveillance. The pixel-based method has clear foreground detection result, high detection precision, low background modeling speed and large memory cost, can provide precise foreground detection result for the subsequent operation of application. Each of the two kinds has its advantages, so this paper focuses on complex scenes, puts emphasis on the division of processing unit and make a deep study on background modeling and foreground detection.Aimed at the region-based method, this paper proposes a background modeling and foreground detection method based on ViBe and block histogram. Under the framework of ViBe, the image is divided into fixed size blocks, histogram is used to describe the block, the current block histogram and the neighborhood block histogram are used for the background modeling. We use MDPA distance as the measurement of the distance between block histograms and update the background database using the max model strategy. In the experiment, we choose three groups of frames from I2R dataset to carry out the contrast experiments with GMM and ViBe. The experiment results show that the proposed method is faster and has higher precision copying with the complex scenes which contain strong disturbance.Aimed at the pixel-based method, this paper proposes a background modeling and foreground detection method based on SLIC superpixels. The image is divided by SLIC superpixels, Gaussian model is set up for each superpixel respectively and a minimum standard deviation is set for all models. We use the background modelings of the current superpixel and the neighborhood superpixels to detect the current pixel and combine the conservative update with the Gaussian model propagation to update the background. During the experiment, three groups of frames from the public dataset are used to carry out the contrast experiments with GMM and the original ViBe. The experiment results show that the proposed method is more outstanding dealing with the complex scenes.