In this work a method is presented for robust estimation of the respiratory frequency from exercise ECGs. The method is based on the rotation angles of the heart's electrical axis as induced by respiration, determined by aligning successive QRS loops to a reference loop using a least squares criterion. The respiratory frequency is estimated by power spectral analysis of the estimated rotation angle series. Special attention has been paid to handling highly non-stationary and noisy exercise ECGs. The ECG and respiratory signals of 14 volunteers and 20 patients were simultaneously recorded during stress testing to evaluate the method. The respiratory frequency was estimated from exercise ECGs and compared to the frequency obtained from respiratory signals. An error of 6.1%±3.7% (0.024±0.017 Hz, mean±sd) is achieved, suggesting that the method is useful for analysis of exercise ECGs.