An Adaptive Bandpass Filter Based on Temporal Spectrogram Analysis for Photoplethysmography Imaging

Photoplethysmography Imaging (PPGI) in the sense of remote vital sign measurement via camera has attracted high interest in recent years. The non-contact measurement principle allows the use in many health monitoring applications, like monitoring of newborns. Beyond that, there are interesting areas of application in the multimedia sector, such as measuring the reaction to multimedia content or heart rate based liveness detection for multimedia security. The derived signal of a PPGI algorithm is often referred as blood volume pulse signal (BVP). The signal corresponds to the optical signal of blood volume changes in the upper skin layers. Most current approaches use peak detection in frequency spectrum to estimate heart rate from BVP signals. However, we focus on heart rate computation based on beat-to-beat peak detection in time domain. In this paper, we present a method for adaptive bandpass filtering for PPGI based on temporal spectrogram analysis of the BVP signal with a sliding time window. The main goal of this new method is to further improve accuracy of beat-to-beat peak detection in time domain. The approach exploits the analysis of main frequency components of the BVP signal over time, to build a bandpass filter with adaptive cutoff frequencies in order to filter noise and interference. So far, state-of-the-art approaches have usually used fixed cut-off frequencies in the physiologically possible range of heart rate. The novelty of the proposed method lies in its simple but effective solution to reduce the influence of noise and interference in the PPGI signal to improve peak detection for heart rate estimation. We show the improvements applying the adaptive bandpass filter technique to four basic algorithmic approaches of PPGI, namely ICA, Chrominance, POS and 2SR and comparing against current state-of-the-art peak detection approaches. For the evaluation we used a database with videos of 26 subjects in 4 different scenarios, each lasting two minutes.