A behvior recognition approach is proposed based on time-frequency analysis and machine learning techniques to identify Parkinson's disease (PD) patients' behviors using local field potential (LFP) signals obtained from a deep brain stimulation (DBS) system. Specifically, the amplitude-time-frequency-variance features are extracted by the matching pursuit decomposition (MPD) algorithm from LFP signals sampled by a DBS lead from the subthalamic (STN) area. Using the extracted feature vectors, different hidden Markov models (HMMs) including discrete and continuous HMMs are trained and then used to recognize different human behviors. The experiment results demonstrate the feasibility, effectiveness and accuracy of our proposed method.