Abstract: Aiming at the problem of "dimension disaster" in human behavior recognition and the classification algorithm has low recognition accuracy and data sample set. First, the SPLDA algorithm is used to obtain the most important principal components with the original sample covariance matrix unchanged, and multiple projection vectors are obtained by greedy search method. Then, the optimal transformation matrix is obtained by updating the class inner divergence matrix. Finally, the dimensionally reduced sample data set is identified by the XGBoost classifier. Experimental results show that compared with the random forest algorithm, the average recognition accuracy of XGBoost classifier is improved by 2.66% and the recognition time is reduced by 0.52s. SPLDA-XGB algorithm can achieve effective dimensionality reduction and has higher accuracy rate of human behavior recognition than PCA algorithm, LDA algorithm, LPP algorithm, l-pca algorithm combined with XGBoost classifier.