Reconstruction of human action is currently the blind spot of image processing problems. For its characteristics of problem, this paper proposed a method based on LLE and KRR. Firstly, it used the image of color and depth information to extract the human skeleton, and established training set of action. Secondly, it turned the training set into action vectors library, and evaluated the human action of low-dimensional manifold according to LLE algorithm. Thereafter, it analyzed the data of manifold. Finally, according to the KRR algorithms, it inverse-mapped the prediction action points of low-dimensional back to higher-dimensional Euclidean space, and reconstructed actions. The experiment results show that the perfect reconstruction actions can be gotten by the mapping of manifold action points, and the proposed method of human action reconstruction has a great effect.