The present pose estimation methods are mainly based on geometry analysis or linear transformation, which are complex and are not universal. A new method is proposed based on nonlinear transformation. According to manifold learning theory, different head poses lie on some low dimensional manifolds. Kernel Principal Component Analysis (KPCA) is a nonlinear dimension reduction method. The hidden manifold in the high dimensional space can be successfully embedded to a low dimensional space using KPCA. A pose curve is gotten using KPCA train samples and new pose image is projected onto this curve. The pose angle can be estimated using interpolation method. The disadvantage of traditional linear method is conquered by KPCA and the experimental result shows that the method is effective to estimate head poses. The method to improve the estimating result is suggested based on the experiments.