Study Of Gaze Tracking Algorithm In Natural Light

Non-contact gaze tracking technologies based on camera obtain image information of eyemovement characteristics though the camera, study and establish the mapping model of theeye gaze direction or the gaze points and then calculate it, which have a great significance asa novel way of human-computer interaction. Most of gaze tracking technologies based oncamera are under the active infrared light, but they have a limited scope of application. Thegaze tracking technologies in natural light without the infrared light can be more widelyapplied. But there is no established theoretical model for them. This article will further studythe gaze tracking technologies in natural light and propose a more perfect algorithm modelwhich includes three aspects:(1)Extract eye feature information in natural light: Because image in natural light has theproblems of low contrast and non-uniform illumination, several methods are adopted asfollowing. Firstly, the eye classifier based on Haar feature and template matching isestablished to locate the eye area. Secondly, the boundary features of iris and the eyelid areacquired based on the maximum between class variance and morphological processing. Atlast, the iris feature parameters are extracted though the improved edge extraction; the innereye corner is located though the corner location operators.(2)Construct the gaze mapping model: Firstly, the dynamic gaze mapping model is built.The eye movement vector composed of the iris center and the inner eye corner is used toestablish the initial static gaze mapping model. The function between the eye movementvector and the iris axis when head moves is established and then combined with the staticgaze mapping model in order to achieve the dynamic gaze mapping model, which has acertain anti-interference ability of head movement. Secondly, the polynomial equation basedon saliency detection is proposed to map the eye movement vector and the gaze point on thebasis of the dynamic gaze mapping model. The non-significant terms of the polynomialequation is eliminated to make the final gaze mapping model more precise.(3)Establish the head movement compensation model: Although the gaze mapping modelconstructed in this article has a certain anti-interference ability of head movement, the gazeestimation error is still large when the head move in a big scope. Therefore, with the long axis, short axis, rotation angle of the iris and the eye movement vector which can represent thehead motion information, the head movement compensation model is established though theRBF neural network training based on the gaze mapping model. It can improve the gazeestimation accuracy in a wide range of head motion case.In this article, the rate of the eye feature information extraction is20frames per second,the iris localization accuracy rate is98%and the inner corner of the eye position error is (1.10.3) pixel. Using the gaze mapping model combined with the head movementcompensation model, the gaze estimation angle is about1degree when the head move in20cm20cm10cm space. The experiment results prove that the gaze tracking algorithmmodel in this article can solve the encountered problems of eye gaze tracking in natural light.