Human behavior recognition has importa nt applications in video surveillance and human-computer interaction,and it is also an important and challenging issue.Behavior recognition using depth informa tion becomes a hot and popular spot.So far,human action recognition based on d epth information has made some achievements,but it still does not have robust descriptors,and their accuracies are still not satisfactory.To solve above problems,this paper proposes three robust action descriptors for depth in formation,and then SVM classifier is adopted.Experiments on two public and challenging behavior recognition datas ets show that our descriptors have strong robustness,distinction and stability,whose performance is much bette r than that of the state of the art algorithms.