| Automatic face tracking and locating is an important and front-line subject in computer vision domain, is the premise of face recognition. Face detection, which is the pre-work of face tracking, has the application in a wide range of fields such as personal identification, visual surveillance. Human-face concerned applications are getting nearer and nearer to practical use. According to the need of the embedded system of omni-direction security authentication and broadcast based on facial feature and RFIC, we have made a deep study on face tracking and locating in complex backgrounds.Omni-direction security authentication system is a hot spot that researched by current domestic and international scholars. The reason called it omni-direction is because this security authentication system has the following three key areas:1. Biometric recognition based on the surface characteristics .The facial characteristics haves some advantages, such as non-contact, remote gathering, extracting and so on, which could achieve a security authentication system of subtle form.2. Application of non-contact RFIC card. RFIC card with 64 internal binary numbers, which is only in the world and determined while manufactured, there are 16 dedicated keys space, which are used to store password set by this system; there are certain storage space used for storage corresponding data information (such as surface-feature). RFIC cards can be written close and read distant, in this way, it could combine with surface-feature information and become the information carrier for the non-contact concealed security authentication system.3. Use of embedded systems. Small size, low power, dedicated hardware, software curable, stable and reliable, using embedded real-time operating system management, professional, become the hardware platform and infrastructure for the system.4. Networking and stand-alone operation are all allowed. Traditional security authentication systems are based on the network, the promise of its role is the network support, once appear a crowded network, intrusion, fault and other conditions, the system will be restricted; but the embedded system of omni-direction security authentication and broadcast based on facial feature and RFIC, we can use the network and also stand-alone operation, suitable for small, medium size, large space, such as various types of security applications. As curing system software, this system could prevent the environmental impact effectively, procure safety management.In sum, such a system has excellent security, is worthy of studying depth.In order to extract the human face in color images accurately, a human face detection and locating algorithm based on skin color was presented. Specific process is as follows.1. The pretreatment is an important part during the process of face detection. As a result of different environment image gathered, such as illumination and equipment performance superior or inferior, input pictures often exist the noise and insufficient in contrast gradient. Moreover, the distance is far or near, the focal distance is big or small and so on, which will cause face is indefinite in the view picture image, such as its size and position. In order to make sure the size, position and the consistent between the face and the image, the pretreatment to the image is necessary. The image pretreatment's main purpose is eliminates the irrelevant information in the image, restores the useful real information, enhance detect abilities of the related information and maximum limit to simplify data, and then enhance the precision of face localization, the track and reliability. First, apply Linear light compensation law to eliminate the effects of uneven illumination; Second,use the multi-frame difference to access regional campaign, and then using binary and morphological processing, which will separated regional campaign from the background for the preparation of face detection. 2. Apply the space in the YCbCr based on the color of face detection and the nonlinear piecewise color transformation, the region of color is accessed. As YCbCr space is linear transformed from RGB space, Cb and Cr, in addition contain color information, also contain a small amount of brightness information. However, experiments show that YCbCr space has separated brightness information well. Cb and Cr component have a good color clustering features, combined with nonlinear transformation, separate color areas. Anil K. Jain, etc. selected the 853,571 color pixels manually from 137 people face image contained in the Heinrich-Hertz-Institute (HHI) image database, and then made a color model in the YCbCr color space. These pixels in the YCbCr projection show a two spindle-shaped tip, meaning that in the area which Y value is larger or smaller, color clustering regional projection in Cb-Cr plane will be reduced, and significantly different from the regional centre. Because YCbCr color format is constrained by linear transformation, the luminance component Y is not completely independent of the color information. In the color space, the color clustering regional clustering has a non-linear change as Y value changed. Consider the effect on the cluster caused by the changes of Y value in the results of the model, it is necessary to amendment the color value under YCbCr color format, before judging the color regional select a sub nonlinear color transformation. Morphology and facial dimensions the proportion of facial dimensions are used to remove the region of non-face.3. Adopt the method of face detection and Kalman filtering, which could predict the locus of the face, to realize the face tracking. The so-called Face Tracking is a process that the face location and size could be detected dynamically in the input image, and then makes sure its trajectory and size changes. Image sequences can be frame sequence taken from the camera, also can be a video streaming. According to the need of omni-direction security certification system, this paper considered the main target tracking technology applications in the face tracking. In the process of tracking, because of shorter time interval between two adjacent images, the small changes of face campaign state, we can assume that people face in the unit time interval is in a uniform movement. Using Kalman filter to estimate the target state in tracking is divided into three stages, namely, filter initialization, state estimation and status updates. Kalman filter could make the mean square error of the estimation error minimized, if the estimation has linear form, then it will be linear minimum variance estimation, that is achieve a optimal state in a certain guidelines.In image sequence, we use Kalman filter to predict face, increased the detection speed and enhanced the accuracy of face detection.The experiment used MATLAB programming to come true,Though affected by background disturb, motion blur, face rotation, face expression change, the experimentation shows the method that this paper adopts is able to track the faces rapidly and effectively, which proves this tracking methods are effective.However, due to the autonomy of the face movement, movement is not necessarily uniform, especially when people face suddenly accelerated, stop or sudden turn, use the method of predicting regional campaign may be have a greater error, during face tracking, we can consider combining means of facial identification to enhance the accuracy of target judgment, can be used for the next step discussion in papers. |
