To improve the rehabilitation effectiveness and reduce the hospital costs, a new upper limb motion recognition model, through which hospital based clinicians can remotely supervise home based stroke rehabilitation, is proposed in this paper. Firstly, the real time limb motion data is collected using a 3-axis accelerometer sensor which is fixed on the upper limb of a patient. Secondly, the Wavelet Transform is employed to extract the approximation coefficients of different types of rehabilitation motions. Finally, a recognition model is established based on an LVQ neural network. 2 typical rehabilitation motions, Bobath handshaking and wrist turning, were chosen to test this proposed recognition system. The experiment results indicate that the recognition accurate rate can achieve as high as 100%. This pilot forms a foundation to further develop a home based remote training and assessment system for stroke rehabilitation.