Due to practical limitations, it is often difficult to measure the responses of civil structure at its all degrees-of-freedom (DOFs). Kalman filter (KF) provides a promising way for state estimation with limited measurements. However, external excitations are required for the implementation of classic KF. Moreover, it is not applicable for nonlinear systems. To circumvent the aforementioned limitations, a KF-based approach is proposed for joint estimation of responses of nonlinear structure and the unknown loading applied to it. By using a projection matrix, a revised observation equation, where the unknown loading is not explicitly presented, is obtained. Based on data fusion of limited acceleration and displacement measurements, the so-called drift problem in the estimated unknown inputs and structural states is avoided. The effectiveness of the proposed approach is numerically validated via a truss structure respectively equipped with magnetorheological (MR) damper. Results show that the proposed approach can satisfactorily estimate the unmeasured nonlinear structural responses and unknown loads in a real time manner.