Respiratory rate (RR) estimation using Photoplethysmogram (PPG) signals has the advantage of high usability and wearability. However PPG sensor is very sensitive to motion artifacts, resulting that the RR features derived from the four respiratory-induced variations (intensity, frequency, amplitude and pulse width) of PPG may present significant inconsistency values. To address this problem, we propose an adaptive fusion approach based on Kalman Filter (KF) to adaptively fuse the RR features in the PPG signals. The model applies the relationship of inter-feature coherence and intra-feature statistical changes to identify the measurement process and the four RR state processes of the KF for the four variations intensity, frequency, amplitude and pulse width, respectively. The fusion of the four estimated RRs from state space of the KF is performed according to the instant Kalman gain and feature consistency metric. The experimental results of 42 subjects show that the proposed adaptive fusion model can effectively improve the estimation accuracy, especially when the four RR features are significantly diverse.