Gaussian Distribution Midel For Video Detection

Video Detection means that people extract their interesting things from a series of video sequences. At present the detection of moving objects from a video sequence has become a hot focus , called Moving Target Detection. Video detection is a core part of the visual analysis of human movement. It will accurately separate the moving objects from video images .That is the the foundation of intelligent control system, the detection and tracking of the human body, and so on. This paper introduces the application and the status quo of Video Detection,and elaborates on the Gaussian distribution model in the application of Video Detection,about both a single Gaussian model and Gaussian Mixture Models for background. With the development of Adaptive Background update and the adaptive learning rate which always has been improved on, the application matures. Finally, there are the Background Subtraction algorithm using Adaptive background mixture models and a improved one in accordance with the characteristics of video detection for traffic flow .The visual analysis of human movement included the video image sequence detection, identification, tracking people and understand their behavior and description. At first this paper introduces its applications for live , study , medical and so on,such as smart surveillance ,perceptual interface ,motion analysis, virtual reality, video search etc. And it elaborate the steps to actualize the visual analysis of human movement which takes video detection as an important part . Video images of moving target detection is extracting a moving target from the continuous video image,by the way of background subtraction,temporal difference,optical flow,level set methods and so on. The purpose of Target Classification is to extract the regional movement of human beings from the detected regional movement. There are two criterions shape-based classification and motion-based classification. The key of Target Classification is the criterion. During the Practical Problem ,we usually deal with the selected target areas or established target by image processing to get a better result. Digital Image Processing is the same as Computer Image Processing. Moving target tracking equivalent to the corresponding match problem on frames in a row to create ,images based on the position,velocity, shape, texture, color and other relevant characteristics. There are several ways of tracking model-based tracking,region-based tracking,active contour based tracking,feature-based tracking and Multi-tracking algorithm. Understanding and description of behaviors mainly face to analysis and recognition of the movement pattern analysis and recognition and description with the natural language . Technology matching time-varying data including DTW (Dynamic Time Warp ing),HMM s (Hidden Markov Models)and NN (Neural Network). At the present there are two ways to identify the actions: template matching and state space approaches.Then the paper gradually introduces Gaussian distribution background model which is one kind of the background difference based on a statistical model of in video detection. Difference Method which equivalent to image subtraction method is a kind of image processing methods which used to detect changes and moving objects, divided into the simple difference method under a controlled environment and difference method based on the background model. In accordance with the characteristics of the background , background model can be divided into single-mode and multi-modal .So Gaussian distribution background model can be divided into single Gaussian distribution background model and multi-Gaussian distribution background model. The former establishes single Gaussian distribution model for the color distribution of each pixel. The latter requires a number of distributions to describe the color distribution of one pixel together,and each gaussian distribution have different weights and priorities. They always rank in accordance with the descending order of priority,and take the appropriate part of weights to the background and thresholds. Only several distributions within the threshold are taken as background,others are foreground. In practice the degree of match is inspected between the value of a pixel and Gaussian distributions in the background model for each pixel. When the value of a pixel matches with the Gaussian distribution ,the pixel is considered as a point of the background,else it is a point of the foreground. Since the establishment of background model is used ,it needs to update the background. Pay attention to the following principles:the response speed of background model to the changes of the background must be fast enough,and background model must have strong anti-jamming capability to moving targets. For the single Gaussian distribution background model and multi-Gaussian distribution background model, update rate and that of the weights are response to the speed of the model updated. Endue the background pixel and the static foreground pixel with higher update rate ,and endue the foreground pixel with lower update rate. That can protect the background model from moving target and respond to the changes of the background at the same time. Obviously multi-Gaussian distribution background model is better than the single Gaussian distribution background model.Finally detail adaptive background mixture models for real-time tracking,Chris Stauffer[10],and a improved one in accordance with the characteristics of video detection for traffic flow. The other make improve on the Background Subtraction algorithm using Adaptive background mixture models:background model matching uses only the luminance information ; ranking Gaussian mixture models according to weight and variance ; selecting background model using monocular depth information ; making dynamic adjustment of sampling rate.