Study On Detection And Tracking Of Human Motion In Visual Surveillance

Intelligent visual surveillance has been receiving increasing attention in the field of computer vision. The purpose of the research on this technology is to make the visual surveillance intelligentized and liberate people from the heavy work. So the research of this field is very useful. In this paper human motion detection and tracking are being studied .The video studied in this paper is obtained from single digital camera.Firstly, the human detection methods of moving people are discussed in the paper. Because there are many noises in the video, we must filter them first. Gaussian lowpass filter is used to smooth the image.The method of background subtraction is used to segment the moving regions . In order to guarantee reliable motion detection, the background image is updated frequently . The small hole in the foreground can be eliminated by morphological filtering. The connected components labeling method is used to find the components which are connected. Then the numbers and areas of the components can be obtained. If it's height, width , and ratio of the height and the width can satisfy the characteristic of human , the component will be considered as a person.Then the Camshift algorithm and the Kalman filtering algorithm are applied to tracking the human motion.Camshift algorithm is continuously adaptive Mean Shift algorithm and it operates on the color probability image. The situation of the face should be obtained fist. In this paper, the initial situation is obtained by seeking the face in the body region which is obtained by background subtraction method. If the area found satisfies the face's characteristic, we consider that the object should be a face. In order to avoid the disturb of hands, the seeking range is located in the top one-fourth of the body. Then tracking of the body motion is studied using Camshift algorithm. This time Camshift algorithm operates on the binary images. Walking and running of human are both tracked using Camshift algorithm. The experiments demonstrate the results achieved effectively.The Kalman filter is a set of mathematical equations that provides an efficient computational means to estimate the state of process, in a way that minimizes the mean of the squared error. Centroid of the body is selected as the feature point for tracking. The experiments demonstrate that the tracking is efficient, even if the person is running and is partially occluded. And we also know from the experiment that Kalman filter algorithm is better than Camshift algorithm when they track the object moving at high speed. But at this time the prediction error of Kalman filtering increases, the object is not entirely detected. A method which combines Kalman filtering and Mean Shift is applied to solve that problem. Experiments show the tracking is efficient .At last, Camshift algorithm and Kalman filter algorithm are used to track two persons. In this paper an improved Camshift algorithm is used to sovle two man's tracking. When overlap of the two people occurs using Camshift algorithm, the seeking range is limited in the top one-fourth of the body and a gridding seeking method is used. When using Kalman filter algorithm, the overlap region is treated as a whole object to track.When the two persons separate from each other after overlap, classifying them is difficult. In order to classify them, the features of the two should be selected. In this paper, the color of the clothes is selected. And the HSV (Hue Saturation Value) color model is used. According to the characteristic of the HSV model, the HSV color space is divided into two part, namely, gray region and hue region. If the color of the clothes is in the gray region, the V value is selected as the feature. And if the color is in the hue region, the feature is H value. Using this method the clothes which have the different color types or have the different gray levels can be distinguished. Through analysis of the relation between the body's moving and corlor, the method of selecting the position where color is extracted is proposed, namely, selecting the color features at different places on the body according to the direction of human's moving. And the method should make the discrimination evident between the different kind of templates, and the features stable in the same kind of templates. The template matching algorithm used in this paper is also improved. When calculating the comparability between the person and the template, we only select the templates in the same moving direction . Otherwise, when calculate the colors distance between the body and the template, we make the far distance of the colors has the heavy weight.After classifying the two persons, the tracking algorithm is initialized according to the new information of them. Then the tracking of two people after overlaping is realized.