Functional near-infrared spectroscopy (fNIRS) is a noninvasive and portable neuroimaging technique that uses NIR light to monitor cerebral activity by the so-called haemodynamic responses (HRs). The measurement is challenging because of the presence of severe physiological noise, such as respiratory and vasomotor waves. In this paper, a novel technique for fNIRS signal denoising and HR estimation is described. The method relies on a joint application of compressed sensing theory principles and Taylor-Fourier modeling of nonstationary spectral components. It operates in the frequency domain and models physiological noise as a linear combination of sinusoidal tones, characterized in terms of frequency, amplitude, and initial phase. Algorithm performance is assessed over both synthetic and experimental data sets, and compared with that of two reference techniques from fNIRS literature.