Emotion identification, which aims to determine a person’s affective state automatically, has immense potential value in many areas, such as action tendency, health care, psychological detection and human-computer (robot) interaction.In this paper, we propose a novel method for identifying emotion from natural walking. After obtaining the three-axis acceleration data of wrist and ankle recorded by smartphone, we run a moving average filter window with different length w, then cut actual data into slices. 114 features are extracted from each slice, and Principal Component Analysis(PCA) is used for feature selection. We train SVM, Decision Tree, Multilayerperception, Random Tree and Random Forest classifiers, and compare the accuracy of emotion identification using different datasets (wrist vs. ankle) on different models. Results show that acceleration data from ankle has better performance in emotion identification than wrist. Among these models, SVM has the highest accuracy of 90.31% for identifying anger vs neutral, 89.76% for happy vs neutral, and 87.10% for anger vs happy. The model for identifying anger/neutral/happy yields the best accuracy of 85%-78%-78%. The results show that we could identify peoples emotional states through the gait of walking with high accuracy.