A method for gazing detection of human eyes is proposed by using Support Vector Machine(SVM) based theoretically on statistic learning theory(SLT).According to the criteria of structural risk minimization of SVM,the errors between sample-data and model-data are minimized and the upper bound of predicting error of the model is also decreased simultaneously so that the ability of generalization of the model is much improved.The simulation results show that when limited training samples are used,the obtained correct recognition rate of the testing samples can be as high as 100% which is much better than some previous results by other methods.The higher processing speed enables the system distinguishes the gazing direction in real-time,as well as to better approach to the characteristics of gazing detection of human vision.