Video pulse wave detection method based on Rayleigh quotient optimization

Abstract

The invention provides a video pulse wave detection method based on Rayleigh quotient optimization. Implicit pulse wave signals (BVP signals) in an RGB video sequence are extracted through Rayleigh quotient optimization. The method comprises the following specific steps: firstly, tracking a region of interest (ROI) through a face detection and feature point extraction technology, and obtaining an original RGB signal through spatial averaging; converting the RGB signal into a YUV signal through matrix transformation, extracting a low-frequency component by using a five-order Butterworth low-pass filter, and subtracting the low-frequency component from the YUV signal to obtain a high-frequency AC signal; then YUV channels are separated, only UV channels are reserved, possible displacement is traversed to construct a cyclic differential matrix, the optimal combination weight corresponding to the minimum characteristic value is found through the Rayleigh quotient theorem, pulse wave signals are extracted through the optimal weight, and high-quality extraction of the BVP signals is achieved.

Description

Technical Field

[0001] The present invention belongs to the field of biomedical signal processing, and specifically is a method for detecting a pulse wave (BVP) from a video using the Rayleigh quotient. The method uses a video acquired by a camera to extract a high-quality pulse wave signal by optimizing the Rayleigh quotient. Background Art

[0002] Pulse wave signals are an important physiological indicator of the human cardiovascular system and are widely used in areas such as heart rate monitoring, blood pressure estimation, and emotion analysis. Traditional methods for acquiring pulse wave signals rely primarily on contact sensors (such as photoelectric pulse sensors). Although these methods have high accuracy, they have the following limitations:

[0003] Invasive: Requires direct contact with the skin and may cause discomfort or skin irritation.

[0004] Poor portability: The device is large in size and not convenient for long-term continuous monitoring.

[0005] High cost: Professional medical equipment is expensive and difficult to popularize.

[0006] In recent years, remote photoplethysmography (rPPG) has garnered widespread attention as a non-contact pulse wave signal extraction technology. rPPG extracts pulse wave signals related to the heartbeat by analyzing subtle color changes in the skin area of ​​a video. However, existing rPPG methods face challenges in practical applications, such as ambient lighting variations, camera noise, and motion artifacts, making it difficult to guarantee the quality of the resulting BVP signal. To address this issue, this paper proposes an rPPG algorithm based on Rayleigh quotient optimization, aiming to efficiently extract BVP signals while ensuring quality. Summary of the Invention

[0007] This invention proposes a video pulse wave detection method based on Rayleigh quotient optimization. This method utilizes color space conversion and Rayleigh quotient optimization to detect pulse wave signals. This method combines low-pass filtering to extract chrominance signals, constructs a cyclic difference matrix, and solves for the minimum Rayleigh quotient to obtain optimal weights to acquire a high-quality BVP signal. The video pulse wave detection method comprises the following steps:

[0008] The video of the facial skin area is captured by the camera and the RGB original signal within the ROI is obtained by spatial averaging, and the RGB signal is converted into a YUV signal through color space conversion.

[0009] The UV channel signal in YUV is retained, the low-frequency components are extracted separately through low-pass filtering, and the low-frequency components are subtracted from the original signal to obtain the UV AC signal.

[0010] The number of bits of the cyclic shift is determined by the video frame rate and heart rate frequency range. The cyclic difference matrix is ​​constructed by traversing the interval, and the eigenvector corresponding to the minimum value of the Rayleigh quotient in this process is updated. After the traversal is completed, the optimal weight of the UV channel is obtained.

[0011] Apply the optimal weight to perform linear combination on the UV AC signal to obtain the final pulse wave signal. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] Figure 1 This is a framework diagram of the video pulse wave detection method based on Rayleigh quotient optimization.

[0013] Figure 2 This is the signal flow chart during the method processing.

[0014] Figure 3 The Ruili optimization flowchart for finding the optimal weight. DETAILED DESCRIPTION

[0015] Combine Figure 1 The present invention detects pulse waves by processing video signals, primarily involving the following steps: spatial averaging, color space conversion, UV channel preservation and filtering, constructing a cyclic difference matrix and determining optimal weights through Rayli optimization, and then combining the UV high-frequency AC signals using these weights to obtain the pulse wave. The principles and specific formulas for each step are as follows:

[0016] Step 1: Suppress the influence of quantization noise on the pulse wave by spatially averaging the pixels in the ROI area in the video frame.

[0017] Step 1.1: Given a video frame sequence containing a face, each frame consists of pixels with coordinates i and j. At time t, the pixels can be represented by vector c. i,j (t)=(r i,j (t),g i,j (t),b i,j (t)) T To describe, that is, the transposed red r i,j (t), green g i,j (t) and blue b i,j (t) channel. The time-varying original RGB signal is obtained by averaging the facial skin pixels in all frames, that is:

[0018]

[0019] Where N is the total number of pixels in the ROI, is the RGB mean at frame time t, and the obtained RGB time domain signal is as follows Figure 2 .a shown.

[0020] Step 1.2, convert the RGB signal into a YUV signal and remove the brightness channel that does not contain pulse wave information. This conversion is linear and can be achieved through matrix transformation, namely:

[0021]

[0022] Then the entire RGB time-varying sequence is converted into a YUV time-varying sequence and can be expressed as:

[0023]

[0024] The pulse wave information mainly exists in the chromaticity. Converting the signal from RGB to YUV space can separate the brightness and chromaticity, and finally retain the chromaticity signal, which is convenient for extracting a cleaner pulse wave. The UV chromaticity signal is as follows: Figure 2 .b shown.

[0025] Step 2: Low-pass filter the UV channel signal to extract the DC component and low-frequency trend, and subtract the filtered signal from the original signal to obtain the UV AC component. Here, a Butterworth filter is used, and its transfer function is:

[0026]

[0027] Where n is the filter order. In this invention, we use a fifth-order Butterworth filter with a cutoff frequency of f c Set to 0.6Hz. Perform low-pass filtering on the UV channel time series signal:

[0028]

[0029] Where * represents the convolution operation, and h(t) is the time domain impulse response of the Butterworth filter H(f). Subtract the low-pass filtered signal from the original UV signal:

[0030]

[0031] The UV high frequency AC signal obtained is as follows Figure 2 .c

[0032] Step 3: Construct a cyclic difference matrix and use the Rayleigh quotient optimization to find the optimal weight. The optimization process is as follows: Figure 3 shown.

[0033] Step 3.1, combine the UV signals processed in step 2, namely:

[0034] X(t)=[U h (t)V h (t)]

[0035] For a signal vector y of length N, a cyclic shift of T bits can be achieved by multiplying it by an N×N cyclic shift matrix P T When T=1, P T It can be defined as:

[0036]

[0037] Then the cyclic shift X(t) is shift (t) can be expressed as:

[0038] X shift (t) = P T X(t)

[0039] For simplicity, X(t) is abbreviated as X, and the combination weight of the UV channel is ω. Then, apply ω to perform a linear combination of X to obtain y:

[0040] y=Xω

[0041] The pulse wave signal is quasi-periodic. The difference between the cyclically shifted signal and the original signal can be minimized at certain moments. The ω that can minimize the difference is the combination weight we want. Let the optimization problem be:

[0042]

[0043] The purpose of limiting the weight of ω here is to avoid the situation where the weight is too large or too small. The original objective function can be expanded as:

[0044]

[0045] Let A = X T (IP T ) T (IP T )X, then the original formula can be transformed into:

[0046]

[0047] According to the definition of Ruili quotient,

[0048]

[0049] Where A is an n×n symmetric matrix, x is a non-zero n×1 vector, x T is the transpose of x. For the size of the Rayleigh quotient, it has the following properties:

[0050] λ min ≤R(A,x)≤λ max

[0051] For our objective function, it is obvious that A is a real symmetric matrix, then the original formula meets the definition of Rayleigh quotient, and at this time ω T ω=1, according to the properties of the Rayleigh quotient, the original problem is transformed into finding the minimum eigenvalue of the matrix A, which can be achieved through eigenvalue decomposition:

[0052] (T * ,ω * )=argmin T,ω λ min (A(T,ω))

[0053] where λ min is the minimum eigenvalue of A. T here needs to be determined according to the video frame rate and heart rate frequency range. Taking the video frame rate of 30 Hz and the heart rate range of 0.6 Hz to 3 Hz as an example, the traversal range of T can be 10 to 48.

[0054] Step 4: After obtaining the optimal weight, perform linear combination on X to obtain Figure 2 The final pulse wave signal shown in .d is:

[0055] y=Xω opt 。

Claims

1. A video pulse wave detection method based on Rayleigh quotient optimization, the main features of which include: S1. Extract the original RGB signal from the ROI area through spatial averaging and convert it into a YUV signal, retaining only the UV signal; S2. low-pass filter the UV signal to obtain a low-frequency component, and subtract the low-frequency component from the YUV signal to obtain a high-frequency AC signal; S3. Construct a cyclic difference matrix based on the phase shift and determine the optimal weight using Rayleigh's theorem; S4. Combine the UV signals using the optimal weights to obtain a pulse wave signal.

2. The pulse wave signal detection method according to claim 1, wherein the UV signal is extracted from the ROI in S1, For each frame of the video sequence, the RGB channels of all pixels within the ROI are averaged to obtain the RGB three-way signal of the corresponding time frame. The RGB signal is converted into a YUV signal using matrix transformation, and finally only the UV chrominance signal is retained.

3. The pulse wave signal detection method according to claim 1, wherein the AC component is extracted from the converted UV signal in S2, characterized in that: A fifth-order Butterworth filter is applied to the two UV signals for low-pass filtering. The cutoff frequency of the filter is 0.6 Hz, which is used to extract the low-frequency and DC signals. The signal obtained after low-pass filtering is subtracted from the UV signal to obtain the UV AC signal.

4. The pulse wave signal detection method according to claim 1, wherein the optimal weight is determined by using the Rayleigh theorem in S3, The number of bits of the cyclic shift is determined according to the camera frame rate and heart rate frequency range. The corresponding cyclic difference matrix is ​​constructed by traversing within this interval. The eigenvector corresponding to the minimum eigenvalue of the Rayleigh quotient is obtained. The eigenvector with the minimum corresponding eigenvalue is selected from all eigenvectors as the optimal combination weight. 5 . The pulse wave signal detection method according to claim 1 , wherein the processed UV signals are linearly combined using optimal weights in S4 to obtain a one-dimensional pulse wave signal.

6. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the program, the method for optimizing the detection of pulse wave signals using the Rayleigh quotient as described in any one of claims 1 to 5 is implemented.

Images