The classical ISOMAP (isometric feature mapping) method is prone to suffer from the noise and the size of neighborhood. A novel method called “Incremental ISOMAP” based on locally estimated geodesic distance for dimensionality reduction was presented. First, this method assumed that the neighborhood of a point located at the highly twisted placed of the manifold might not be linear so that its neighbors should be determined by geodesic distance. Then, incremental learning was used to replace the batch mode in pattern recognition,  aiming to enhance the ability of real time. The proposed method is simple, general and easy to deal with high-dimensional data. The experimental results on face recognition show that the method is efficient and robust.