An improved action recognition method is proposed based on the signals acquired by a smart phone acceleration sensor to reduce the complexity of the traditional action recognition method and enhance the recognition rate. The blind selection method is applied in feature extraction stage, which means using principal component analysis (PCA) method to reduce dimensionality and eliminate multi-dimensional interference, while the selected features have no corresponding physical significance. In classification and identification, the genetic algorithm is used to optimize support vector machine (SVM) classifier. Experimental results indicate that the proposed method can accurately recognize actions such as walking, standing, running and climbing stairs.