Research On Method Of Precise Localization Of The Pupil Based On Video Images

As one of the key research subjects in the field of computer vision and patternrecognition, precise localization of the pupil is closely related to many disciplinessuch as pattern recognition, image processing, intelligent information processing,neural network and computer vision. Originally, it has gradually developed along withadvancement of technologies such as face detection and facial expression recognition.Actually, the development of the pupil detection technology is based on exploration offace recognition beginning in the1970s. At present, as face and iris technologies havemade rapid progress, the pupil localization technology is required to have a moresuitable and better way to match. This paper proposes a new pupil localizationalgorithm on the basis of existing research results, which describes the followingresearch contents:1. A face detection algorithm based on skin color is adopted in this paper, whichcan detect faces using special nature of the skin color. This is because the skin colorhas these features: it shows important information of the face; it is independent ofsubtle facial characteristics; it is suitable for rotation and illumination variation; it hashigh stability. First, RGB image was converted to YCrCb image with good clusteringcharacteristics, then the elliptical model for skin color was used to detect the skinpixels, and then some noise and skin color-like regions were removed through themorphological operation, and finally faces were detected accurately through the facialgeometric features.2. In this paper, the eyebrow and eye region is extracted from face image throughthe Sobel operator. As some facial features in this region such as eyes, eyebrows andnose show more distinct horizontal edge information, the changed convolution kernelcan be used to process original images, which enables the horizontal edge of face tobe more centralized and clear. In addition, the eyebrow and eye region can besegmented more efficiently through close combination and efficient coordination ofits two features: it has a certain width and it is in the line with a larger horizontal projection value.3. This paper proposes a new pupil localization algorithm on the basis oftraditional localization algorithm. First, some candidate points from the radius of theeyebrow and eye region in the pupil circle within the range of the human eye weredetected based on the principle of Hough transformation; then the pupil circle wasscreened out through making using of two features: the grayscale value is the lowestand there exists the biggest variation in the pupil circle; finally, as “sift” is notsensitive to some features such as rotation, zoom and horizontal movement, the “sift”algorithm was adopted to obtain the pupil point—the point with the largest scale inthe pupil circle. In addition, as the upper half of the face has been asymmetric, whichis caused by the fringe and the human’s looking around with the eyes, the lower halfof the face was segmented out first and then the perpendicular bisector of the face wasextracted through making use of symmetry between “sift” and human face in somefeatures (scale, direction and angle); finally, the purpose of determining the gazingdirection of the human eye was achieved based on distance. As the algorithm in thispaper is designed to extract local features of the human eye, it maintains invariance torotation, brightness, etc.; meanwhile, it has inhibiting effect on noise to some extent.The method introduced in this paper has the features of cumulative progression andtight link in steps, which enables the characteristics of “sift” to have a full play,ensuring high accuracy and fast computing of this location system.