To avert catastrophic failure in the structures, joints are typically designed to yield, but not fail, so that energy accumulated under cyclic loading is dissipated. Eventually, this renders the structural joints to be characteristically weaker and more vulnerable than the members. Yet, damage detection research mostly assumes damage in the members only. This article proposes a model-based predictor-corrector algorithm that uses an interacting filtering approach to efficiently estimate joint damage in the presence of input and measurement uncertainties. For the predictor model, a novel strain-displacement relationship specific to semi-rigid frames is developed to map nodal displacements to corresponding strain measurements. The proposed estimation method embeds robustness against non-stationary input (e.g. seismic excitation) in the state filter, itself. For this, an output injection technique is integrated within the state filter. The modified state filter (robust Kalman filter) runs within an enveloping parameter filter (Particle filter) to simultaneously estimate the system states and joint damage parameters, respectively, using the response signal. Strain has been adopted as measurement since it is frame independent (beneficial for seismic activity) and also comparatively cheaper to use. Numerical studies are performed on a two-dimensional three storythree bay shear frame for different joint damage locations and severities. The sensitivity and the stability of the proposed approach are further investigated. Experimental validation of the proposed algorithm is carried out on a 2-D steel frame.