The indirect extraction of respiratory frequency during exercise testing is very interesting and challenging. In this work we propose a method to estimate respiratory frequency during exercise testing from heart rate variability (HRV) analysis. Empirical mode decomposition is first applied to HRV signal to obtain the intrinsic mode functions (IMF). The combination of different IMF is studied and a criterion is proposed to select the one which best represents respiratory information. Finally, time-frequency analysis is applied to the combination of IMFs and respiratory frequency is selected as the largest peak of the spectrum within a restricted frequency band and given that the spectrum is sufficiently peaked. The proposed methodology is applied to a database consisting of the instantaneous RR interval series of 23 healthy and sportive volunteers recorded during treadmill exercise testing. Estimated respiratory frequency shows a relative error close to 3% with respect to the respiratory frequency simultaneously recorded by a gas analyzer system.