Objective. Efficient non-contact heart rate (HR) measurement from facial video has received much attention in health monitoring. Past methods relied on prior knowledge and an unproven hypothesis to extract remote photoplethysmography (rPPG) signals, e.g. manually designed regions of interest (ROIs) and the skin reflection model.Approach. This paper presents a short-time end to end HR estimation framework based on facial features and temporal relationships of video frames. In the proposed method, a deep 3D multi-scale network with cross-layer residual structure is designed to construct an autoencoder and extract robust rPPG features. Then, a spatial-temporal fusion mechanism is proposed to help the network focus on features related to rPPG signals. Both shallow and fused 3D spatial-temporal features are distilled to suppress redundant information in the complex environment. Finally, a data augmentation strategy is presented to solve the problem of uneven distribution of HR in existing datasets.Main results. The experimental results on four face-rPPG datasets show that our method overperforms the state-of-the-art methods and requires fewer video frames. Compared with the previous best results, the proposed method improves the root mean square error (RMSE) by 5.9%, 3.4% and 21.4% on the OBF dataset (intra-test), COHFACE dataset (intra-test) and UBFC dataset (cross-test), respectively.Significance. Our method achieves good results on diverse datasets (i.e. highly compressed video, low-resolution and illumination variation), demonstrating that our method can extract stable rPPG signals in short time.
Keywords: fusion attention mechanism; heart rate estimation; remote photoplethysmography; short-time monitoring; spatial-temporal convolution.
© 2022 Institute of Physics and Engineering in Medicine.