To overcome deficiency of low recognition efficiency and easy to be interferenced by environment of existing methods for human action recognition

a new human action recognition method based on mixed joint feature is proposed. Mixed joints feature is made up of different skeleton descriptors

including kinetic energy

potential energy

joint orientation and joint angle

which makes the feature representation more effective. Long short-term memory（ LSTM） neural network is used as classifier to recognize human action

which has better ability to handle time series. Experiments on MSR Action 3D dataset show that the mixed joints feature performs better than many other existing skeletal features.