This paper proposes a new algorithm for human action recognition based on variation energy images features, it is defined as VEI maps. Firstly, a temporal-spatial curve is defined using ratio of width to height of a moving human body, and it is used to determine a starting point and an ending point of a human action so that an accurate action fragment can be extracted from videos. Secondly, VEI is calculated based on this action fragment, and eight dimensions of features are defined using this VEI map, in which there are four dimensions of shape features and four dimensions of motion features which are based on Haar-like methods. Finally, eight features are used to train an SVM to distinguish human actions. The experimental results show that the average accuracy is better than other approaches.