A new algorithm for eye location is presented. In this method, the probable eyes candidate regions are localized rapidly by using the combination of the variance filter and the independent analysis (ICA) method. The method depends on higher-order statistics of the training images decorrelated by basis vectors. Compared with principal component analysis (PCA) and traditional method of independent component analysis, the method does better in the eye identification. The results of experiments show that the proposed method has the detection rate of 97.3%, and is more efficient and robust to the variable postures and illuminations.