We propose a general method for online estimation of the quality of movement from Kinect skeleton data. A robust non-linear manifold learning technique is used to reduce the dimensionality of the noisy skeleton data. Then, a statistical model of normal movement is built from observations of healthy subjects, and the level of matching of new observations with this model is computed on a frame-by-frame basis following Markovian assumptions. The proposed method is validated on the assessment of gait on stairs.