Principal component analysis of photoplethysmography signals for improved gesture recognition

In recent years, researchers have begun to introduce photoplethysmography (PPG) signal into the field of gesture recognition to achieve human-computer interaction on wearable device. Unlike the signals used for traditional neural interface such as electromyography (EMG) and electroencephalograph (EEG), PPG signals are readily available in current commercial wearable devices, which makes it possible to realize practical gesture-based human-computer interaction applications. In the process of gesture execution, the signal collected by PPG sensor usually contains a lot of noise irrelevant to gesture pattern and not conducive to gesture recognition. Toward improving gesture recognition performance based on PPG signals, the main contribution of this study is that it explores the feasibility of using principal component analysis (PCA) decomposition algorithm to separate gesture pattern-related signals from noise, and then proposes a PPG signal processing scheme based on normalization and reconstruction of principal components. For 14 wrist and finger-related gestures, PPG data of three wavelengths of light (green, red, and infrared) are collected from 14 subjects in four motion states (sitting, walking, jogging, and running). The gesture recognition is carried out with Support Vector Machine (SVM) classifier and K-Nearest Neighbor (KNN) classifier. The experimental results verify that PCA decomposition can effectively separate gesture-pattern-related signals from irrelevant noise, and the proposed PCA-based PPG processing scheme can improve the average accuracies of gesture recognition by 2.35∼9.19%. In particular, the improvement is found to be more evident for finger-related (improved by 6.25∼12.13%) than wrist-related gestures (improved by 1.93∼5.25%). This study provides a novel idea for implementing high-precision PPG gesture recognition technology.