Remote Blood Pressure Estimation via the Spatiotemporal Mapping of Facial Videos

Figures

• 

  Figure 1. An overview of our blood pressure estimation network. We first generate the corresponding spatiotemporal feature map slices of the input facial video clips. Then, a feature extractor composed of a depth residual network and BiLSTM fits a high-dimensional feature, which is fed into the blood pressure classifier to locate the blood pressure interval through this feature. Finally, the blood pressure calculator combines the results of the feature extractor and the blood pressure classifier to output the final blood pressure value.

• 

  Figure 2. An illustration of the generation of spatiotemporal feature map slices from an input facial video clip of T frames. Firstly, we use Seetaface6 to detect the human face, localize 68 landmarks, and define four ROIs; then, we calculate the average pixel values of each ROI to get spatiotemporal maps. Secondly, the data augmentation module randomly masks a part of the cropped spatiotemporal maps presented above along both the time domain and the spatial domain; then, they are transformed from the RGB color space into the modified YUV color space. Finally, the cropped video sequence is sliced to generate the spatiotemporal feature map slices for the following network.

• 

  Figure 3. The distributions of the ground-truth blood pressure values in the MPM-BP dataset.

• 

  Figure 4. Devices and setup used to collect MPM-BP.