Non-contact in-vehicle person blood pressure detection method

Abstract

The invention provides a non-contact in-vehicle person blood pressure detection method, which comprises the following steps: acquiring video data of in-vehicle persons according to a vehicle-mounted image acquisition module, and acquiring a video frame sequence containing a face region; positioning a face region in the video frame by adopting a face detection algorithm, and cutting a face image from the video frame; preprocessing the cut face image, including image normalization and de-noising operation, to obtain a preprocessed face image; a time sequence model is adopted to analyze dynamic changes among the multiple video frames, and time sequence characteristic signals related to blood pressure fluctuation are extracted; calculating systolic pressure and diastolic pressure values of the target object through a deep learning algorithm according to the extracted spatial feature signal and time sequence feature signal; and transmitting the calculated blood pressure index data to a cloud server for storage through the wireless communication module.

Description

Technical Field

[0001] The present invention relates to the field of information technology, and particularly to a non-contact method for detecting the blood pressure of vehicle occupants. Background Art

[0002] Implementing non-contact blood pressure monitoring in a vehicle environment faces many challenges. First, the light conditions inside the vehicle are complex and variable, which may lead to unstable face detection and feature extraction. Second, the human body posture changes frequently during driving, increasing the difficulty of extracting blood pressure-related feature signals. Third, factors such as vehicle vibration and acceleration / deceleration introduce noise, affecting the accuracy of blood pressure estimation. In addition, drivers with different skin colors, ages, and genders may exhibit different facial features, requiring the model to have strong generalization ability. At the same time, there is a high real-time requirement, and the entire process from video acquisition to blood pressure estimation needs to be completed quickly with limited computing resources. On the other hand, long-term driving may cause fatigue, which in turn affects the facial blood flow characteristics. How to ensure the reliability of blood pressure estimation in this situation is a major problem. Finally, considering driving safety, the system needs to complete the monitoring without disturbing the driver's attention, which poses higher requirements for human-computer interaction design. How to build a robust, efficient, and accurate in-vehicle non-contact blood pressure monitoring system under these constraints is a complex and challenging technical issue. Summary of the Invention

[0003] The present invention provides a non-contact method for detecting the blood pressure of vehicle occupants, mainly including: According to the video data of vehicle occupants collected by the in-vehicle image acquisition module, obtain a video frame sequence containing the face region; Use a face detection algorithm to locate the face region in the video frame and crop out the face image from the video frame; Preprocess the cropped face image, including image normalization and denoising operations, to obtain the preprocessed face image; Input the preprocessed face image into a pre-trained convolutional neural network model, and extract spatial feature signals related to blood pressure through multi-layer convolution and pooling operations; Use a time series model to analyze the dynamic changes between multiple video frames and extract time series feature signals related to blood pressure fluctuations; According to the extracted spatial feature signals and time series feature signals, calculate the systolic and diastolic blood pressure values of the target object through a deep learning algorithm; Transmit the calculated blood pressure index data to the cloud server for storage through a wireless communication module; The cloud analysis system combines the historical blood pressure data of the target object to analyze the current blood pressure index and generates personalized health suggestions including warning information and lifestyle recommendations.

[0004] The technical solutions provided by the embodiments of the present invention may include the following beneficial effects: Description of the Drawings

[0005] Figure 1 It is a flowchart of a non-contact blood pressure detection method for in-vehicle personnel according to the present invention.

[0006] Figure 2 It is a schematic diagram of a non-contact blood pressure detection method for in-vehicle personnel according to the present invention.

[0007] Figure 3 It is another schematic diagram of a non-contact blood pressure detection method for in-vehicle personnel according to the present invention. Detailed Embodiments

[0008] To further understand the content of the present invention, the present invention will be described in detail in combination with the drawings and embodiments. The present application will be further described in detail below with reference to the drawings and embodiments. It can be understood that the specific embodiments described herein are only used to explain the related invention, rather than limiting the invention. In addition, it should be noted that for the sake of description, only the parts related to the invention are shown in the drawings.

[0009] Such as Figures 1-3 , a non-contact blood pressure detection method for in-vehicle personnel in this embodiment may specifically include: S101. According to the vehicle-mounted image acquisition module, collect the video data of the in-vehicle personnel, and obtain a video frame sequence including the face region.

[0010] Obtain the video data through the vehicle-mounted module, capture the dynamic images of the in-vehicle personnel by using a camera to obtain the original video sequence. Extract each frame image from the original video sequence, use the inter-frame difference method to detect the moving area, and determine the frame sequence including the personnel. For the frame sequence including the personnel, apply a face detection algorithm to locate the face area to obtain a face localization frame. If multiple areas are detected in the face localization frame, judge the main face area through geometric features to obtain an optimized face frame. According to the optimized face frame, extract a continuous video frame sequence to obtain the time series data of the face region. Analyze the inter-frame changes through the time series data, use the optical flow method to track the dynamic features of the face region, and determine the motion trajectory. For the motion trajectory, combine video analysis techniques to generate a complete video frame sequence including the face region.

[0011] Specifically, the in-vehicle image acquisition module collects video data at a rate of 30 frames per second through cameras deployed inside the vehicle, compresses the video stream using H.264 encoding to reduce storage requirements, and sets the resolution to 1280×720 to ensure the clarity of the face region. In the video frame processing stage, first, a deep learning model based on YOLOv5 is used for face detection. This model is pre-trained on the COCO dataset and fine-tuned for the in-vehicle scenario, and the detection confidence threshold is set to 7 to ensure stable recognition under varying lighting conditions (e.g., within the range of 100 - 1000 lux). For the detected face region, the MTCNN algorithm is used for key point localization, and a coordinate matrix containing 5 feature points (the centers of the left and right eyes, the tip of the nose, and the corners of the mouth) is output. Subsequently, the face is aligned to the standard 112×112 size through affine transformation. In the feature extraction link, the ArcFace model is used to generate a 512-dimensional feature vector. This model is trained on the VGGFace2 dataset, and the cross-entropy loss function and an additional angular margin penalty (margin = 5) are adopted. In the feature comparison stage, cosine similarity is calculated, and the threshold is set to 4 to distinguish different occupants. For consecutive video frames, the Kalman filter algorithm is used to track the face movement trajectory, and the IOU threshold between the predicted box and the detection box is set to 6. When the target is lost for 5 consecutive frames, the tracking is determined to terminate. The system synchronously records the timestamp (with a millisecond-level accuracy) and the vehicle CAN bus data (such as vehicle speed, steering angle), and analyzes the driver's attention state through multi-modal data fusion. For example, when the vehicle speed exceeds 60 km / h and the eyes-closed frequency > 3 Hz is detected continuously for 2 seconds, a fatigue warning is triggered. All processing is completed on the embedded platform Jetson Xavier NX, and the inference delay is controlled within 50 ms to meet the real-time requirements.

[0012] S102. Use a face detection algorithm to locate the face region in the video frame and crop out the face image from the video frame.

[0013] Obtain a video frame through video input, extract single-frame data from the frame sequence, process it using a face detection algorithm to obtain the face region coordinates. For the face region coordinates, determine the region boundary, and generate a first face image through a cropping operation. If there is noise in the first face image, use a denoising algorithm to process it to obtain a second face image. Based on the second face image, detect the face feature points, judge whether the number of feature points reaches the preset threshold to determine the image integrity. Through the integrity judgment result, obtain the qualified face image, and use image extraction technology to save the third face image. For the third face image, analyze the detection accuracy to obtain the final face image data. Extract key information from the final face image data to generate a structured output result.

[0014] Specifically, when locating the face region in a video frame, the MTCNN (Multi-task Cascaded Convolutional Networks) algorithm based on deep learning can be adopted. MTCNN precisely locates the face step by step through three cascaded neural networks (P-Net, R-Net, and O-Net). First, P-Net conducts a preliminary detection on the input video frame to generate candidate face regions. Assuming the resolution of the input video frame is 1920×1080, P-Net will output multiple candidate boxes, and the size of each box is approximately 50×50 pixels. Then, R-Net further filters these candidate boxes, removes the misdetected boxes, and retains the boxes with a confidence level higher than 7. Finally, O-Net finely adjusts the remaining candidate boxes, outputs the final face region, and generates the coordinates of five key points (such as eyes, nose, and corners of the mouth). Assuming the size of the final face box output by O-Net is 120×120 pixels and it is located at the position (800, 400) in the video frame. Next, according to the coordinate information output by O-Net, the face image is cropped from the video frame. When cropping, the cropping area can be appropriately enlarged, for example, expanding the 120×120 pixel box to 150×150 pixels to ensure that complete face information is included. The cropped face image can be further used for tasks such as face recognition or expression analysis.

[0015] For example, the cropped face image is input into the FaceNet model to extract a 128-dimensional feature vector for subsequent identity verification or clustering analysis. The entire process is implemented through an automated script without manual intervention to ensure efficiency and accuracy.

[0016] S103. Preprocess the cropped face image, including image normalization and denoising operations, to obtain the preprocessed face image.

[0017] The face image is separated from the original data through the cropping operation to obtain the initial face image. The preprocessing is performed on the initial face image. The normalization technique is used to adjust the brightness range to obtain the brightness-balanced image. The denoising operation is applied to the brightness-balanced image to smooth the image noise through Gaussian filtering to obtain the denoised image. If the noise in the denoised image still exceeds the preset threshold, the denoising operation is repeated to obtain the optimized image. According to the characteristics of the optimized image, it is judged whether the brightness balance condition is satisfied to obtain the judged image. Edge detection is performed on the judged image, and the Sobel operator is used to extract the edge information to obtain the edge-enhanced image. The edge-enhanced image is obtained and the format conversion is performed to output the preprocessed face image.

[0018] Specifically, for face image preprocessing, normalization is first performed. The input image is uniformly adjusted to 112×112 pixels using the bilinear interpolation algorithm to ensure the consistency of subsequent processing dimensions. At the same time, the mean-variance normalization method is used to standardize the pixel values with the mean [485, 456, 406] and variance [229, 224, 225] of the ImageNet dataset. The calculation formula is (x - mean) / std, so that the pixel values are distributed in the interval [-1, 1]. For the problem of uneven illumination, the CLAHE algorithm of histogram equalization is applied, the cropping grid is set to 8×8, and the contrast threshold is limited to 0. The feature contrast is enhanced by adaptively adjusting the gray distribution of the local area. In the denoising stage, the non-local means denoising algorithm is used, the search window is set to 21×21 pixels, the similar block window is 7×7, and the filtering parameter h = 10. The weighted average value is calculated using the principle of image self-similarity to eliminate Gaussian noise. For the remaining high-frequency noise, wavelet threshold denoising is used. The sym4 wavelet basis is selected for 3-layer decomposition, and the soft threshold processing is performed on the detail coefficients. The threshold calculation formula is σ√(2logN), where σ is obtained by the median estimation method to calculate the noise standard deviation, and N is the total number of wavelet coefficients. Finally, through the edge enhancement algorithm, the Sobel operator is used to calculate the gradient magnitude, the convolution kernel size is set to 3×3, and the pixels with gradient values exceeding the threshold of 15 are enhanced by 2 times to improve the clarity of the facial feature contours. All processing is performed in the YCbCr color space, and only the transformation is applied to the luminance component Y channel to avoid chromaticity information distortion.

[0019] S104. Input the preprocessed face image into a pre-trained convolutional neural network model, and extract the spatial feature signals related to blood pressure through multi-layer convolution and pooling operations.

[0020] Process the preprocessed face image through a pre-trained convolutional neural network model, and extract spatial feature signals using multi-layer convolution and pooling operations. Obtain the preprocessed data from the face image and input it into the network model for preliminary feature calculation. Use a convolutional neural network to perform multi-layer convolution operations to obtain an initial spatial feature representation. Process the spatial features through pooling operations to extract refined feature signals. If there is noise in the feature signals, filter it using a preset threshold to determine the signal components related to blood pressure. According to the filtered feature signals, use a support vector machine classifier to judge the blood pressure status category. Compare the classification result with the preset blood pressure range to obtain the final blood pressure-related information.

[0021] Specifically, the preprocessed face image is input into a pre-trained convolutional neural network model. First, through the first convolutional operation, 32 3×3 convolutional kernels with a stride of 1 are used to extract features from the input image, resulting in 32 feature maps. Then, a 2×2 max pooling layer with a stride of 2 is used to downsample the feature maps, reducing the computational amount while retaining the main features. The second convolutional operation uses 64 3×3 convolutional kernels with a stride of 1 to further extract more complex features, generating 64 feature maps, and also performs downsampling through a 2×2 max pooling layer. The third convolutional operation uses 128 3×3 convolutional kernels with a stride of 1 to extract higher-level features, generating 128 feature maps, and again performs downsampling through a 2×2 max pooling layer. After the convolutional and pooling operations, the feature maps are flattened into a one-dimensional vector and input into the fully connected layer. The ReLU activation function is used for non-linear transformation. The fully connected layer contains 256 neurons, which are used to further extract the spatial feature signals related to blood pressure. Finally, through the output layer using the Sigmoid activation function, the feature signals are mapped to blood pressure values. The output layer contains 1 neuron, representing the predicted blood pressure value. In the whole process, the Adam optimization algorithm is used, and the learning rate is set to 0.01. The model parameters are updated through the backpropagation algorithm to minimize the mean square error between the predicted value and the true value, and finally an accurate blood pressure prediction result is obtained.

[0022] S105. Analyze the dynamic changes between multiple video frames using a time series model, and extract the time series feature signals related to blood pressure fluctuations.

[0023] Analyze the video frames through a preset time series model, extract the inter-frame dynamic changes, and obtain the feature data. Use a convolutional neural network to process the feature data, obtain the change trend, and determine the time series signal. Perform filtering processing on the time series signal to remove noise and obtain a smooth signal. If the smooth signal exceeds the preset threshold, analyze the frequency components through Fourier transform to judge the fluctuation period. Compare the fluctuation period with the blood pressure fluctuation database to obtain the correlation coefficient and determine the fluctuation-related features. Adjust the time series model parameters through the correlation coefficient to obtain the optimized feature extraction result. Update the signal processing flow using the optimized feature extraction result to obtain the final time series signal.

[0024] Specifically, in the time series model analysis, first, feature extraction is performed on video frames through a convolutional neural network (CNN). For example, the ResNet-50 model is used to perform convolutional operations on each frame of the image to extract a 128-dimensional feature vector. Then, these feature vectors are input into a long short-term memory network (LSTM), and the hidden layer of the LSTM is set to 256 units to capture the dynamic changes between frames. Through the output of the LSTM, a time series feature signal can be obtained, which reflects the temporal dependence between video frames. To extract features related to blood pressure fluctuations, principal component analysis (PCA) can be used to reduce the dimensionality of the time series feature signal, retaining the first 10 principal components, which can explain more than 90% of the variance. Then, a support vector machine (SVM) is used to classify the features after dimensionality reduction. The kernel function of the SVM is selected as the radial basis function (RBF), and the parameters C and γ are optimized through grid search, and finally, the classification accuracy rate is 85%. To further analyze the specific patterns of blood pressure fluctuations, the dynamic time warping (DTW) algorithm can be used to calculate the similarity of blood pressure fluctuations in different time periods. For example, within a 10-second time window, fluctuations with a DTW distance less than 5 are considered similar. Finally, through clustering analysis (such as the K-means algorithm), similar blood pressure fluctuation patterns are classified, and the value of K is set to 3 to obtain three main blood pressure fluctuation patterns, which can be used for subsequent blood pressure prediction and health monitoring.

[0025] S106. According to the extracted spatial feature signal and time series feature signal, calculate the systolic blood pressure and diastolic blood pressure values of the target object through a deep learning algorithm.

[0026] Obtain the original signal data of the target object through a sensor, and use signal extraction technology to separate the spatial feature signal and the time series feature signal. Based on the separated spatial feature signal and time series feature signal, construct a feature-based data set. Use a deep learning algorithm to train the feature-based data set to obtain a systolic blood pressure prediction model and a diastolic blood pressure prediction model. Through the trained systolic blood pressure prediction model, calculate the systolic blood pressure value of the target object. Through the trained diastolic blood pressure prediction model, calculate the diastolic blood pressure value of the target object. If the calculated systolic blood pressure value and diastolic blood pressure value exceed the preset threshold, adjust the feature signal extraction parameters through object analysis. According to the adjusted feature signal, repeat the above deep learning calculation process to obtain the final systolic blood pressure and diastolic blood pressure values.

[0027] Specifically, first, the pulse wave signal at the wrist of the target object is collected by a photoelectric sensor. The sampling frequency is set to 125 Hz, and 30 seconds of raw data is continuously collected. The raw signal is preprocessed. A Butterworth band-pass filter (5 - 8 Hz) is used to remove baseline drift and high-frequency noise. The filtered signal extracts spatial features through wavelet transform. The db4 wavelet is used for 5-layer decomposition to obtain the detail coefficients D1 - D5 and the approximation coefficient A5. Among them, the energy value of the D4 band (2 - 4 Hz) is calculated as 7 millivolt squared as the main spatial feature. For temporal feature extraction, the sliding window analysis method is adopted. The window width is set to 256 sampling points (about 2 seconds), and the step size is 64 points. The standard deviation of the peak interval of the pulse wave within each window is calculated as 18 milliseconds as the variability index. The extracted 18-dimensional spatial features and 15-dimensional temporal features are input into the designed dual-channel convolutional neural network. The spatial feature channel uses 3 layers of one-dimensional convolution (kernel sizes are 5, 3, 3 respectively, and the number of channels is 16 / 32 / 64), and the temporal feature channel uses a bidirectional LSTM layer (64 hidden units). The features of the two channels are fused in the fully connected layer and then the predicted values of systolic blood pressure and diastolic blood pressure are output. The network is trained using the MIMIC-III clinical dataset, adopting the mean squared error loss function and the Adam optimizer (learning rate 0.01). After 200 rounds of training, the average absolute error of the test set reaches 3 mmHg (systolic blood pressure) and 8 mmHg (diastolic blood pressure). When the final model is deployed, after performing the same feature extraction and normalization processing on the real-time collected pulse wave signal, the trained network is input and the blood pressure prediction value can be output. For example, when the spatial feature vector [7, 2,..., 8] and the temporal feature vector [18, 6,..., 12] are obtained in a certain measurement, the model outputs a systolic blood pressure of 126 mmHg and a diastolic blood pressure of 82 mmHg.

[0028] S107. Transmit the calculated blood pressure index data to the cloud server for storage through the wireless communication module.

[0029] Obtain the calculated blood pressure index data through the wireless communication module, and determine the transmission format using a preset communication method. Extract the index data from the above transmission format, and judge whether the data is completely received through data transmission. If the data is completely received, send the index data to the cloud server and obtain the server-side confirmation signal. Judge the storage data status according to the confirmation signal, and determine whether the storage is successful using a preset threshold. Obtain the stored data through the server-side, and obtain a structured result using data processing technology. Extract the transmission process record from the structured result, and judge the stability of the communication method through the random forest algorithm. Optimize the wireless communication parameters according to the stability judgment result to obtain the adjusted communication method.

[0030] Specifically, in a blood pressure monitoring device, first, blood pressure data of the user is collected by a sensor. For example, the systolic blood pressure is 120 mmHg, the diastolic blood pressure is 80 mmHg, and the heart rate is 75 beats per minute. These data are preliminarily processed by a built-in microprocessor, and the Kalman filtering algorithm is used to filter the noise to ensure the accuracy of the data. The processed data is transmitted to the cloud server at a speed of 10 times per second through a wireless communication module, such as Bluetooth or Wi-Fi. During the transmission process, the AES-256 encryption algorithm is used to encrypt the data to ensure the security of the data. After receiving the data, the cloud server uses machine learning algorithms to analyze the data. For example, through the support vector machine (SVM) model, the blood pressure trend of the user in the next 24 hours is predicted, and the analysis results are stored in a distributed database, such as MongoDB, for subsequent query and analysis. At the same time, the server will automatically trigger an alarm system according to a preset threshold, such as the systolic blood pressure exceeding 140 mmHg or the diastolic blood pressure exceeding 90 mmHg, and send notifications to the user or medical staff. Throughout the process, the system will record the timestamp and status of each data transmission to ensure the integrity and traceability of the data.

[0031] S108. The cloud analysis system combines the historical blood pressure data of the target object to analyze the current blood pressure indicators, and generates personalized health suggestions including warning information and lifestyle suggestions.

[0032] Obtain historical data and current indicators through the data uploaded by the target object. Extract the change trend of blood pressure indicators from the historical data, and conduct a comparative analysis in combination with the current indicators. If the change trend exceeds the preset threshold, determine the category of warning information through the analysis process. Obtain the corresponding content of life suggestions according to the category of warning information. Integrate the life suggestions and warning information using a pre-established template to obtain a personalized plan. Optimize the personalized plan through machine learning algorithms and judge the integrity of the health plan. Generate the final output content for the optimized health plan.

[0033] Specifically, the cloud analysis system first collects the user's blood pressure data for 30 consecutive days through IoT devices, including daily morning resting blood pressure (systolic pressure 120-145mmHg, diastolic pressure 75-95mmHg) and nighttime blood pressure fluctuations (±8-12mmHg), and uses the time series database InfluxDB to store and establish a blood pressure trend map with granularity per minute. The system uses a sliding window algorithm (window size 7 days, step length 1 day) to calculate the dynamic threshold, and triggers a level 1 warning when the systolic blood pressure exceeds 140mmHg for 3 consecutive days. The random forest model (including 50 decision trees, maximum depth 10) was used to analyze the influencing factors, and it was found that the Pearson correlation coefficient between the sudden drop in blood pressure at night >15% and caffeine intake (daily >300mg) was 73. The system automatically generates a recommended plan: control coffee intake to less than 200mg per day, and recommends completing caffeine intake before 4 pm in combination with dynamic blood pressure monitoring data. For users with abnormal morning blood pressure (systolic blood pressure rises > 20 mmHg within 2 hours after waking up), the system calls the LSTM neural network (hidden layer 128 units, dropout rate 2) to predict the risk in the next 7 days. When the predicted value exceeds 135 / 85 mmHg, morning exercise suggestions are pushed, and a 30-minute brisk walking exercise with a heart rate maintained at 100-120 beats / minute is recommended. All analysis results are transmitted to the mobile terminal in JSON format, and a line chart is used to visualize the blood pressure trend in the past 30 days, and abnormal data points that exceed the WHO standard value are marked in red.

[0034] It should be noted that the above examples are only some specific embodiments of the present invention. Obviously, the present invention is not limited to the above embodiments, and there are many variations. All variations that can be directly derived or associated with the content disclosed by a person skilled in the art should be considered as the protection scope of the present invention.

Claims

1. A non-contact method for detecting the blood pressure of vehicle occupants, characterized in that, The method includes: According to the in-vehicle image acquisition module, acquiring video data of the in-vehicle personnel, and obtaining a video frame sequence containing a face region; Using a face detection algorithm to locate the face region in the video frame, and cropping out a face image from the video frame; Preprocessing the cropped face image, including image normalization and denoising operations, to obtain a preprocessed face image; Inputting the preprocessed face image into a pre-trained convolutional neural network model, and extracting spatial feature signals related to blood pressure through multi-layer convolution and pooling operations; Using a time series model to analyze the dynamic changes between multiple video frames, and extracting time series feature signals related to blood pressure fluctuations; According to the extracted spatial feature signals and time series feature signals, calculating the systolic and diastolic blood pressure values of the target object through a deep learning algorithm; Transmitting the calculated blood pressure index data to a cloud server for storage through a wireless communication module; The cloud analysis system combines the historical blood pressure data of the target object to analyze the current blood pressure index, and generates personalized health advice including warning information and lifestyle suggestions.

2. The method according to claim 1, characterized in that, The step of, according to the in-vehicle image acquisition module, acquiring video data of the in-vehicle personnel, and obtaining a video frame sequence containing a face region, includes: Obtaining video data through the in-vehicle module, using a camera to capture the dynamic images of the in-vehicle personnel, and obtaining an original video sequence; Extracting each frame image from the original video sequence, using the inter-frame difference method to detect the moving region, and determining a frame sequence containing personnel; For the frame sequence containing personnel, applying a face detection algorithm to locate the face region, and obtaining a face localization frame; If multiple regions are detected in the face localization frame, then judging the main face region through geometric features, and obtaining an optimized face frame; According to the optimized face frame, extracting a continuous video frame sequence, and obtaining time series data of the face region; Analyzing the inter-frame changes through time series data analysis, using the optical flow method to track the dynamic features of the face region, and determining the motion trajectory; For the motion trajectory, combining video analysis techniques, and generating a complete video frame sequence containing the face region.

3. The method according to claim 1, wherein The step of, using a face detection algorithm to locate the face region in the video frame, and cropping out a face image from the video frame, includes: Obtaining a video frame through video input, extracting single-frame data from the frame sequence, and processing it with a face detection algorithm to obtain face region coordinates; For the face region coordinates, determining the region boundary, and generating a first face image through a cropping operation; If there is noise in the first face image, then processing it with a denoising algorithm to obtain a second face image; According to the second face image, detecting face feature points, judging whether the number of feature points reaches a preset threshold, and determining the image integrity; Based on the integrity judgment result, obtaining a qualified face image, and using image extraction technology to save a third face image; For the third face image, analyzing the detection accuracy, and obtaining the final face image data; Extracting key information from the final face image data, and generating a structured output result.

4. The method according to claim 1, wherein The step of, preprocessing the cropped face image, including image normalization and denoising operations, to obtain a preprocessed face image, includes: Separating the face image from the original data through a cropping operation, and obtaining an initial face image; Preprocess the initial face image, adjust the brightness range using normalization technology to obtain a brightness - balanced image; Apply a denoising operation to the brightness - balanced image, smooth the image noise through Gaussian filtering to obtain a denoised image; If the noise in the denoised image still exceeds the preset threshold, repeat the denoising operation to obtain an optimized image; According to the characteristics of the optimized image, determine whether the brightness - balance condition is met to obtain a judged image; Perform edge detection on the judged image, use the Sobel operator to extract edge information to obtain an edge - enhanced image; Obtain the edge - enhanced image and perform format conversion to output the pre - processed face image.

5. The method according to claim 1, characterized in that, Input the pre - processed face image into a pre - trained convolutional neural network model, and extract spatial feature signals related to blood pressure through multi - layer convolution and pooling operations, including: Process the pre - processed face image through a pre - trained convolutional neural network model, and extract spatial feature signals using multi - layer convolution and pooling operations; Obtain the pre - processed data from the face image and input it into the network model for preliminary feature calculation; Perform multi - layer convolution operations using a convolutional neural network to obtain an initial spatial feature representation; Process the spatial features through pooling operations to extract refined feature signals; If there is noise in the feature signals, filter it using a preset threshold to determine the signal components related to blood pressure; According to the filtered feature signals, use a support vector machine classifier to judge the blood - pressure status category; Compare the classification result with the preset blood - pressure range to obtain the final blood - pressure - related information.

6. The method according to claim 1, wherein Analyze the dynamic changes between multiple video frames using a time - series model to extract time - series feature signals related to blood - pressure fluctuations, including: Analyze the video frames through a preset time - series model, extract the inter - frame dynamic changes to obtain feature data; Process the feature data using a convolutional neural network, obtain the change trend, and determine the time - series signal; Perform filtering processing on the time - series signal to remove noise and obtain a smooth signal; If the smooth signal exceeds the preset threshold, analyze the frequency components through Fourier transform to judge the fluctuation period; Compare the fluctuation period with the blood - pressure fluctuation database to obtain the correlation coefficient and determine the fluctuation - related features; Adjust the time - series model parameters according to the correlation coefficient to obtain an optimized feature - extraction result; Update the signal - processing flow using the optimized feature - extraction result to obtain the final time - series signal.

7. The method according to claim 1, wherein Calculate the systolic and diastolic blood - pressure values of the target object according to the extracted spatial feature signals and time - series feature signals through a deep - learning algorithm, including: Obtain the original signal data of the target object through a sensor, and use signal - extraction technology to separate the spatial feature signals and time - series feature signals; Construct a feature - basis data set based on the separated spatial feature signals and time - series feature signals; Use a deep - learning algorithm to train the feature - basis data set to obtain a systolic - blood - pressure prediction model and a diastolic - blood - pressure prediction model; Calculate the systolic - blood - pressure value of the target object through the trained systolic - blood - pressure prediction model; Calculate the diastolic - blood - pressure value of the target object through the trained diastolic - blood - pressure prediction model; If the calculated systolic blood pressure value and diastolic blood pressure value exceed the preset threshold, the feature signal extraction parameters are adjusted through object analysis; Based on the adjusted feature signal, the above deep learning calculation process is repeated to obtain the final systolic and diastolic blood pressure values.

8. The method according to claim 1, wherein The transmission of the calculated blood pressure index data to the cloud server for storage through the wireless communication module includes: Obtain the calculated blood pressure index data through the wireless communication module and determine the transmission format using the preset communication method; Extract the index data from the above transmission format and judge whether the data is completely received through data transmission; If the data is completely received, send the index data to the cloud server and obtain the server-side confirmation signal; Judge the storage data status according to the confirmation signal and determine whether the storage is successful using the preset threshold; Obtain the stored data through the server-side and obtain the structured result using data processing technology; Extract the transmission process record from the structured result and judge the stability of the communication method through the random forest algorithm; Optimize the wireless communication parameters according to the stability judgment result to obtain the adjusted communication method.

9. The method according to claim 1, characterized in that, The cloud analysis system combines the historical blood pressure data of the target object to analyze the current blood pressure index and generates personalized health suggestions including warning information and lifestyle suggestions, including: Obtain the historical data and current index through the data uploaded by the target object; Extract the change trend of the blood pressure index from the historical data and conduct a comparison analysis in combination with the current index; If the change trend exceeds the preset threshold, determine the category of the warning information through the analysis process; Obtain the corresponding lifestyle advice content according to the category of the warning information; Integrate the lifestyle advice and warning information using the pre-established template to obtain a personalized plan; Optimize the personalized plan through the machine learning algorithm and judge the integrity of the health plan; Generate the final output content for the optimized health plan.

Images