Research On Key Issues Of Video Processing In The Video Surveillance System

As a part of security system, video surveillance system is a comprehensive system with strong safety assurance abilities. The video surveillance system is intuitive, convenient and rich in information, so it is widely used in many scenarios. With the dramatically rapid technologies developments in Computer Networks, Communications and Multimedia, digital video surveillance system is used in every fields of the national economy. There are several outstanding advantages when using video surveillance system:Strong real-time performance in monitoring targets, good long-distance transmission ability, and high controllable for administrative staffs, so the studies on video surveillance system are strong significant both in Theory and reality.The research topics about video surveillance are quite rich, including target movement detecting and tracking, object classification and behavior understanding. The research involves core open issues in fields of computer vision, pattern recognition, artificial intelligence, and so on. It is a challenging research topic. This thesis proceeds from the view of video processing, aiming at the problems existed in video noise reduction, video stabilization, and target tracking in videos, and the following achievements are obtained.Firstly, the video denoise algorithm is studied. It is different from common image denoising. Video is not only one image, but a combination of frames of images. The images are relevant in time, but the noise is unrelated to time. This makes it possible to filter the video sequence in the time domain to reduce the differences between adjacent frames, and results in a dramatically reduction of screen flicker when people watching videos. The common algorithm is implemented using block as the basic unit, and this brings Blocking Effects. A combined algorithm using Nonlocal Mean and Time Domain Average Filtering is proposed to reduce noise of single image frame and also overcome video flicker appearance.Secondly, video stabilization algorithm is studied. Not every surveillance camera is mounted in fixed position, take vehicle mounted surveillance camera as an example, the recorded video picture will be shaky. A video stabilization algorithm based on particle filtering is proposed. A semi-iterative unscented Kalman particle filter is proposed in advanced, and the main idea of the proposed filter is to give the suggested particle distribution using extended Kalman filter, and then implement particle re-extract particles using unscented Kalman Filter. Aiming at the low computation speed of unscented Kalman filter, scale factor is introduced, the re-extraction of particles is only happened when the estimation error appears to be big. The experimental results show that the estimating accuracy of the proposed algorithm performs better than all the12compared algorithms.The video stabilization algorithm is presented in Chapter5. Video stabilization algorithm generally consists of movement estimation, movement filtering, movement compensation, and image compensation. Comparing to other kinds of video stabilization algorithms, the video stabilization algorithm based on feature points is more robust. But the accuracy of Feature points based video stabilization algorithm is affected by the errors in localizing, selecting and matching of feature points. To solve these problems, a particle filter based video stabilization algorithm is proposed. Firstly, it extracts and matches the feature points of two adjacent frames based on SRUF operator, and then removes local movement vectors using RANSAC algorithm. At the beginning of the movement estimation calculation, the initial movement estimation is acquired from feature points matching data by using least square method; secondly, the accurate movement estimation result is obtained using particle filter algorithm based on the initial estimation. The experiments prove the validity of the proposed algorithm.Target tracking is another important task of surveillance system, and Camshift is a common video tracking algorithm. There is a contradiction problem between windows size expansion and noise interference resistance in Camshift algorithm. The proposed ACshift algorithm is based on the adaptive windowed2D kernel density template Camshift algorithm. It dynamically expands the windows of Camshift algorithm by introducing the windows2D kernel density template, and suppress the effect caused by noises by weighting the pixels in the expanded windows. In the end, it models and estimates the movement trail of the target by using Kalman filter to enhance the tracking ability to the rapid moving objects of ACshift algorithm, and experiments is performed to analyze the problems caused by large proportion covering and similar color interference. The results show that, the improved ACshift algorithm can effectively mediate the contradiction between windows size expansion and noise interference resistance, while just sacrificing a few computing performance. The experiments also prove the improvement to ACshift by using Kalman filter.