It is crucial for walking assistive exoskeletons to detect and comply with the motion intention of the wearers. Recently, the first author has proposed to estimate the wearer's joint torque for intention detection, and conducted assistive walking successfully. The aforementioned torque estimation algorithm was based on the model of the exoskeleton with fixed parameters. However, the model parameters depend on the weight and height of the wearer. In order to achieve more accurate torque estimation, regardless of the physique of the wearer, this paper aims to simultaneously estimate the wearer's joint torque and model parameters. Since the joint torque is periodic during walking, we on-line identify the fundamental frequency of the gait and parameterize the wearer's joint torque by its first two harmonics. Then the parameters of the joint torque and the model are estimated together by the least mean square (LMS) algorithm. Experiments are conducted to verify that the method of this paper achieves robust accuracy in torque estimation, even though additional loads are attached to the exoskeleton. The fitness of the estimated and actual torques are more than 80% in all experimental conditions.