This paper presents a real-time Kinect-based hand pose estimation method. Different from model-based and appearance-based, our approach retrieves continuous hand motion parameters in real time. First, the hand re-gion is segmented from the depth image. Second, some specific feature points on the hand are located by random forest classifier, and the relative displacements of these feature points are converted as a rotation invariant fea-ture vector. Finally, the system retrieves the hand joint parameters by applying the regression functions on the feature vectors. The experimental results are compared with the ground truth obtained by data glove for reliabili-ty evaluation. The effects of different distances and dif-ferent rotation angles to the estimation accuracy are eva-luated. Finally, we ask some other subjects to evaluate the system adaptively of our system to different hand shapes.