A series elastic actuator (SEA) is mainly used in human-robot interaction applications. Especially, a reaction-force-sensing SEA (RFSEA), where the spring is located between the ground and the actuator, has been developed as a practical implementation of the SEAs, as their form-factors are superior to the conventional SEAs. However, the RFSEA has limitations on its torque control performance. The output torque of an RFSEA is estimated as the spring torque, assuming the load is fixed. However, since the load moves in human-robot interaction applications, the assumption causes significant degradation in the control performance of the RFSEA. This paper presents a precise torque estimation method for the RFSEA. The dynamics of the RFSEA were analyzed, modeling the actuator as a two-mass system. The output torque was verified as the function of two variables from the dynamic analysis: the stator angle and the angular velocity of the rotor. In this paper, the torque estimator with a single encoder is proposed to maintain the form-factor of the RFSEA. An encoder measures the stator angle, and a designed observer based on a disturbance observer estimates the angular velocity of the rotor. The performance of the proposed torque estimator was verified experimentally. From the experimental results, an error of the proposed torque estimator was reduced by 37% compared to the conventional when the load moves by mimicking the human movement.