Autoregressive modelling is used to identify periodic physiological signals such as heart rate or breathing rate in an image of a subject. The colour channels of a video signal are windowed and normalised by dividing each signal by its mean. The ratios of the normalised channels to each other are found and principal component analyses conducted on the ratio signals. The most periodic of the principal components is selected and autoregressive models of one or more different orders are fitted to the selected component. Poles of the fitted autoregressive models of different orders are taken and pure sinusoids corresponding to the frequency of each pole are generated and their cross-correlation with the original component is found. Whichever pole corresponds to the sinusoid with the maximum cross-correlation is selected as the best estimate of the frequency of periodic physiological information in the original video signal. The method may be used in a patient monitor or in a webcam-enabled device such as a tablet computer or smart phone.
