In recent years, recognition of human motion on visible light video sequences has made some progress. Since the data sources are sensitive to target color, light intensity and background clutters, the depth information is applied to human motion recognition. In this paper, a local representation method of human movement based on Spatio-Temporal Interest Points(STIPs) is adopted, and the applications of Harris and Gabor filter detection methods on depth information are achieved. A novel Depth Cuboid Similarity Feature(DCSF) is built to describe the corresponding results. Finally, action classification is completed by Support Vector Machine(SVM) classifier based on spatio-temporal codebook. Experimental results demonstrate that detection method of Gabor filter obtains better recognition performance in depth datasets.