We introduce a human arm movement tracking system that has been developed to aid the rehabilitation of stroke patients. A wearable 3-axis inertial sensor is used to capture arm movements in 3-D space and in real time. The tracking algorithm is based on a kinematical model that considers the upper and lower forearm. To improve accuracy and consistency, a weighted least square filtering strategy is adopted. The calculated motion trajectory was compared with that measured using a commerically available Qualysis tracking system. For 3D cyclical rotation, the mean wrist position error was 2.45 cm without filtering and 1.79 cm after the filtering alogorithm was applied. The experimental results demonstrate the favorable performance of the proposed framework in estimation of upper limb motion in stroke rehabilitation.