Monitoring the subject's physical activity intensity (PAI) is key to preventing sports injury and enhancing conditioning performance. Existing monitoring equipment can be cumbersome and is not always accessible in daily situations due to cost or usability. For ease of usage, we propose a motion robust remote photoplethysmography (rPPG) measurement during exercise for contactless PAI detection. To account for vigorous motion during physical exercise, we leverage signal holding with motion analysis to maintain the consistency of the signal source. We then obtain robust rPPG signals via a convolutional neural network (CNN)-based rPPG constructor module composed of a dual feature extractor with signal-based targets, a base span module, and a noise elimination module. In terms of the PAI detection application, as changes in heart rate are a dominant biomarker during exercise, this is applied to estimate the training level for PAI detection. In this study, training status is classified into three levels and is estimated via a residual-learning-based architecture with sequential heart rate input. In terms of heart rate measurements, the proposed method outperforms the state-of-the-art approaches in motion scenarios on both ECG-Fitness and our own exercise datasets. The mean absolute error (MAE) is reduced to about one-third of the comparison methods in real-world treadmill running scenarios, and the error level goes down by about 30% in the ECG-Fitness dataset under a cross-dataset evaluation. In addition, for follow-up PAI detection applications, the heart rate values estimated with the proposed method approximate the heart rate ground truth. It is capable to represent the trend between training load and PAI changes. We propose a novel framework for the motion robust rPPG measurement during exercise applied to PAI detection, with which overall athlete assessment can be accomplished as a convenient alternative to existing athlete monitoring equipment.
|IEEE Transactions on Instrumentation and Measurement
|Published - 2023
All Science Journal Classification (ASJC) codes
- Electrical and Electronic Engineering