Abstract: To improve the accuracy of human action recognition algorithm,a novel approach for action recognition based on composite spatio-temporal features is proposed,which combines 3D histograms of oriented gradients feature with histograms of optical flow feature.Firstly,the composite spatiotemporal features are used to describe the pixels distribution and pixels variance in 3D spatio-temporal local area.Then the composite spatio-temporal feature dictionary is built and used to describe the behavior sequence feature vector.Lastly,the topic model is used to construct the human action recognition algorithm that classifies the composite features extracted from the behavior sequence,which leads to the action recognition.The experimental results show that the proposed algorithm improves the accuracy of human action recognition effectively.