Non-Contact Respiration Measurement Method Based on RGB Camera Using 1D Convolutional Neural Networks

Figure 1. Overall process of the proposed method.

Figure 2. Structure of the ROI detection model. The variance in pixels is the input and the probability that the variance is related to respiration is the output.

Figure 3. Labeling process for video using the reference and the variance in some pixels: (a) sample of a downsampled video; (b) correlation coefficient calculated from the video and reference signal; (c) label for the video (T = 0.7); (d) reference signal; (e–g) variance of Pixels 1–3, respectively.

Figure 4. Visualization of the actual ROI pixels and clustering results of PCA (top) and the variance in some selected pixels (bottom): (a–h) arbitrary points on a two-dimensional plane and the trajectory of the points (pixels).

Figure 5. Clustering results of (a) original data and (b) data including both the original data and symmetric data (dotted circles indicate symmetric data).

Figure 6. Bpm change in guidelines provided to subjects to acquire training data: (a) sequential change in breathing rate; (b) rapid change in breathing rate; (c) patterns involving apnea.

Figure 7. Examples of augmented data used for training.

Figure 8. Data samples from Experiment 2 with background noise.

Figure 9. Bland–Altman plot and regression analysis results for bpm: (a) overall data; (b) Experiment 1; (c) Experiment 2; (d) Experiment 3.

Figure 10. Examples of bpm estimation results and peak detection results in Experiments 1 and 3: (a) change in bpm measurement; (b) examples of peaks detected in the actual signal.

Figure 11. Comparison of signals in Experiment 2.

Figure 12. Measurement results of the proposed method for the case where the apnea section is included during breathing.