Face recognition has higher performance with controlled illumination and pose. But in some applications such as video surveillance, imaging condition is uncontrolled and the subject is not cooperative. In this paper pose invariant face recognition in complex backgrounds is discussed and a framework is proposed. Our algorithm is comprised of four parts. In the first part a face location algorithm combining face feature and template is proposed to determine the face location, represented as center of eyes and mouth. In the second part a face segmentation algorithm using curve fitting is proposed to segment face region in the image. The third part is face normalization---to obtain a front view face from a face with variant pose. In the forth part, the face recognition based on normalized faces is implemented using eigenface method. The algorithm is tested using 70 images of 14 persons, the experimental results confirm the efficiency of our algorithms.