CN114998229B - A non-contact sleep monitoring method based on deep learning and multi-parameter fusion - Google Patents

Abstract

The invention relates to a non-contact sleep monitoring system based on deep learning and multi-parameter fusion, and belongs to the fields of image processing and deep learning. The system firstly segments an acquired sleep video image, then builds a deep convolutional neural network for extracting and amplifying physiological signals, amplifies heart rate signals of a forehead area and eye movement frequencies of an eye area by setting different amplification factors of the network to obtain the forehead area video image of the amplified heart rate signals and the eye area video image after the eye movement frequencies are amplified, and then extracts corresponding frequency spectrums by utilizing fast Fourier transformation to find frequencies corresponding to frequency spectrum peaks as monitored heart rate signals and eye movement frequencies. For a three-position body video image, a sleeping gesture monitoring neural network structure based on deep learning is built, sleeping gesture features extracted from the network are input into a full-connection layer to be subjected to six classification, classification results correspond to six sleeping gestures of supine, prone, left straight lying, left spiral lying, right straight lying and right spiral lying, and the number of turning times is counted through switching of different sleeping gestures in the six sleeping gestures. And finally, comprehensively evaluating the sleep quality by combining the monitored physiological signals. The invention has the characteristics of high comfort, multi-parameter fusion and high automation, and realizes the monitoring of physiological parameters such as heart rate, respiratory rate, eye movement frequency, sleeping posture, turning-over times and the like in a non-contact manner.

Description

A non-contact sleep monitoring method based on deep learning and multi-parameter fusion

Technical Field

The present invention belongs to the field of image processing and deep learning, and specifically relates to a non-contact sleep monitoring system that combines video image processing technology with a deep convolutional network to achieve multi-parameter fusion.

technical background

During sleep, a series of functions of the human brain, muscles, eyes, heart, breathing, etc. will change. Monitoring these changes can help to judge the quality of human sleep. Sleep disorders usually refer to abnormal sleep quantity or quality or certain clinical symptoms during sleep, such as reduced or excessive sleep, sleep breathing disorders, rapid eye movement sleep behavior disorder, etc. Medically proven, people with long-term sleep disorders will induce a variety of diseases, so timely diagnosis and treatment of sleep disorders are of great significance to human health.

Polysomnography is known as the golden method for diagnosing and treating sleep disorders. It mainly monitors the physiological signals of the patient's electroencephalogram, electrocardiogram, electrooculogram, oral and nasal airflow, blood oxygen saturation and other channels, and makes a diagnosis based on the collected signals. When using polysomnography for monitoring, it is necessary to install a variety of sensors on the subject, which brings great discomfort to the subject. In addition, even if it monitors many parameters, the doctor will still use the subject's medical history and subjective feelings during the monitoring period as one of the basis for evaluation, and the interpretation results are highly subjective. With the development of deep learning, many miniaturized sleep monitoring devices have emerged, such as smart pillows, mattresses, and bracelets. Smart pillows and mattresses monitor pressure changes during sleep through pressure sensors and count the number of times the subject turns over. Smart bracelets can monitor heart rate during sleep when worn. Although existing sleep monitoring equipment reduces the discomfort of wearing sensors during monitoring, it can only monitor a single physiological parameter, resulting in inaccurate and incomplete evaluation results of sleep quality.

In response to the above problems in sleep monitoring, we designed a non-contact sleep monitoring system based on deep learning and multi-parameter fusion. For heart rate and eye movement frequency, a deep convolutional neural network is used to extract and amplify tiny physiological signals from sleep videos, while suppressing the generation of artifacts to achieve effective physiological signal amplification. Finally, the amplified physiological signals are subjected to spectrum analysis. For sleeping posture and number of turning over, a convolutional neural network is used to automatically extract the sleeping posture features of the video frames obtained by the three-dimensional camera, and classify them into six categories. The number of turning over during sleep can be monitored by switching between different sleeping postures.

Summary of the invention

In order to solve the problems of discomfort caused by contact monitoring of polysomnography to the subjects and the subjectivity of manual interpretation, as well as the singleness of physiological parameters monitored by other sleep monitoring equipment, the present invention designs a non-contact sleep monitoring system based on deep learning and multi-parameter fusion to achieve non-contact monitoring of multiple physiological parameters such as heart rate, eye movement frequency, sleeping posture, and number of turning over.

The technical solution of the present invention is a non-contact sleep monitoring method based on deep learning and multi-parameter fusion, which includes the following steps:

Step 1: Build a sleep monitoring platform and place three cameras on the upper, left, and right sides of the tester's body to obtain video images of the tester during sleep;

Step 2: Segment the video image obtained by the camera located above the tester's body in step 1 to obtain a video image of the tester's forehead area and a video image of the eye area;

Step 3: Build a deep convolutional neural network for physiological signal extraction and amplification, and use it to extract and amplify tiny physiological signals in the video;

Step 4: Input the forehead area video image and the eye area video image obtained in step 2 into the deep convolutional neural network constructed in step 3, respectively, extract and amplify the heart rate signal of the forehead area and the eye movement signal of the eye area, and output the forehead area video image after amplifying the heart rate signal and the eye area video image after amplifying the eye movement frequency;

Step 5: Separate the RGB three-channels of each frame of the forehead area video image after the amplified heart rate signal obtained in step 4, calculate the average value of the pixels in the three channels of R, G, and B, and then stack them in time series to obtain the pulse wave signal;

Step 6: Perform fast Fourier transform on the pulse wave signal obtained in step 5 to obtain a time series spectrum of the human pulse wave;

Step 7: Perform spectrum analysis on the time series spectrum obtained in step 6, and select the frequency corresponding to the spectrum peak as the heart rate monitoring result;

Step 8: stacking the eye region video images obtained in step 4 after amplifying the eye movement frequency in a time series, and performing a fast Fourier transform to obtain a frequency spectrum of the eye region video images;

Step 9: extracting the frequency corresponding to the spectrum peak of the eye area video image spectrum obtained in step 8 as the monitoring result of the eye movement frequency;

Step 10: Build a deep learning-based sleep posture monitoring neural network structure. If the sleep posture monitoring neural network has not been trained, execute step 11; if the sleep posture monitoring neural network has been trained, execute step 13;

Step 11: Collect more than 1,000 images of the test subject while sleeping by deploying three cameras on the left and right sides of the body, annotate the images with features, and manually annotate the sleeping positions of the test subject, corresponding to six sleeping positions: supine, prone, left side straight, left side curled up, right side straight, and right side curled up;

Step 12: Send the image data annotated in step 11 to the neural network for training, and divide the training set and the validation set into a ratio of 8:2 to train the network until the accuracy of the validation set reaches more than 95%, and the network training is completed;

Step 13: Input the three-dimensional body video images of the tester from the three cameras in step 1 into the trained neural network, and perform six-category output on them, and the classification results correspond to six sleeping positions: supine, prone, left side straight, left side curled up, right side straight, and right side curled up;

Step 14: If the test subject switches between any two of the six sleeping positions in step 13, it is counted as one turn over, but the switch between lying upright on the left side and lying curled up on the left side and the switch between lying upright on the right side and lying curled up on the right side are not counted in the statistics of the number of turns over;

Step 15: Combine the heart rate parameters obtained in step 7, the eye movement frequency obtained in step 9, the sleeping posture obtained in step 13, and the number of tossing and turning obtained in step 14 to comprehensively evaluate the sleep quality results of the tester.

Wherein, the step 2 is specifically as follows:

Step 2.1: For the video image captured by the camera above the tester's body, call the dlib library in Python to segment the face area and extract the tester's face video image;

Step 2.2: Perform facial key point detection on the facial video image obtained in step 2.1, call the dlib library in python to perform facial key point detection, and obtain the positions of 68 key points of the tester's face;

Step 2.3: The video image obtained in step 2.1 is segmented again by using the positions of the key points of the face obtained in step 2.2; a rectangular forehead area video image is segmented by identifying the key points of the center of the left and right eyebrows of the face and the upper boundary of the face identified by the dlib library;

Step 2.4: By identifying the key points of the left and right eyes of the face, find the key points representing the left eye corner, the right eye corner, the uppermost part of the eye socket, and the lowermost part of the eye socket, and segment a rectangular area of the eye area video image through the above four key points;

Wherein, the step 3 is specifically as follows:

Step 3.1: Build the encoder structure of the deep convolutional neural network for physiological signal extraction and amplification. The structure is first: 2 convolutional layers, each of which uses the Relu activation function, followed by 3 residual networks, and then a convolutional layer to extract the physiological signals in the sleep video. The step size is set to 2, and finally the two residual structure outputs are connected;

Step 3.2: Build a modulation amplification structure of a deep convolutional neural network. First, perform a convolution operation on the difference of the physiological signals of the two frames of sleep images through a convolution layer. The activation function is the Relu function, and then multiply it by the amplification factor α. The convolution layer and residual structure are used again to perform nonlinear changes on the amplified features to obtain the amplified physiological signal differential features.

Step 3.3: Build the decoder structure of the deep convolutional neural network, superimpose the amplified physiological signal differential features on the initial sleep image, and then decode and output the amplified video through upsampling and two convolutional layers to achieve the amplification of physiological signals in the sleep video.

Wherein, the step 4 is specifically as follows:

Step 4.1: Input the forehead area video image obtained in step 2 into the deep convolutional neural network built in step 3, set the amplification factor α=15, extract and amplify the heart rate in the forehead area, and output the forehead area video image after amplifying the heart rate signal;

Step 4.2: Input the eye area video image obtained in step 2 into the deep convolutional neural network built in step 3, set the amplification factor α=30, extract and amplify the eye movement frequency of the eye area, and output the eye area video image after amplifying the eye movement frequency;

Wherein, the step 10 is specifically as follows:

Step 10.1: Build a neural network structure, including 4 convolutional layers, 3 maximum pooling layers, 1 fully connected layer and 1 classifier;

Step 10.2: To prevent the amount of calculation from being too large, extract the key frame of the video at the current moment every 1 second, and extract one frame of image from the cameras located on the upper, left, and right sides of the tester's body respectively, forming three-channel image data and inputting them into the neural network in step 10.1;

Step 10.3: The three-channel image data in step 10.2 are respectively passed through a convolution layer with a convolution kernel of 10×10, a 2×2 maximum pooling layer, a convolution layer with a convolution kernel of 10×10, a 2×2 maximum pooling layer, a convolution layer with a convolution kernel of 10×10, a 2×2 maximum pooling layer, and a convolution layer with a convolution kernel of 10×10 to extract image features;

Step 10.4: Input the above extracted features into the fully connected layer for six classifications. The classification results correspond to six sleeping positions: supine, prone, left side straight, left side curled up, right side straight and right side curled up.

Wherein, the step 15 is specifically as follows:

Step 15.1: The heart rate of a person during normal sleep is 60 to 100 beats per minute, and can be reduced to 50 beats per minute during deep sleep. The heart rate of the test subject during sleep is monitored. When the heart rate drops significantly, it is considered that the test subject has entered the deep sleep period, and when the heart rate gradually increases, it is considered that the deep sleep period has been exited. Finally, the proportion of the deep sleep period in the test subject's sleep is calculated. The greater the proportion of the deep sleep period, the higher the sleep quality;

Step 15.2: During the REM sleep stage, the eyeballs will move rapidly. Monitor the eye movements of the test subject. If the eye movement frequency increases significantly, combined with the increase in heart rate in step 14.1, it can be considered that the test subject has entered the REM stage. Count the duration of the REM stage. If the REM sleep is suddenly interrupted, it is often a signal of the onset of diseases such as angina pectoris and asthma.

Step 15.3: Lying flat on the back is considered to be a better sleeping position during sleep, but it is not suitable for people with respiratory diseases or frequent snoring. Instead, sleeping on the side should be adopted; monitor the sleeping position of the tester, and if the tester suffers from respiratory diseases or snores, provide sleeping position adjustment suggestions when they adopt a non-side sleeping position;

Step 15.4: Monitor the number of times the test subject turns over. If the test subject turns over too often, it indicates that the test subject may be calcium ion deficient or mentally stressed, resulting in poor sleep quality.

The present invention is a non-contact sleep monitoring system based on deep learning and multi-parameter fusion. The system first segments the acquired video image, and segments the forehead area video image, the eye area video image, and the three-position body video image. Then, a deep convolutional neural network for physiological signal extraction and amplification is built. By setting different amplification factors of the network, the heart rate signal of the forehead area and the eye movement frequency of the eye area are amplified to obtain the forehead area video image with amplified heart rate signal and the eye area video image after amplifying the eye movement frequency. Then, the corresponding spectrum is extracted by fast Fourier transform, and the frequency corresponding to the spectrum peak is found as the monitored heart rate signal and eye movement frequency. For the three-position body video image, a sleeping posture monitoring neural network structure based on deep learning is built, and the sleeping posture features extracted from the network are input into the fully connected layer for six classifications. The classification results correspond to six sleeping postures: supine, prone, left straight, left curled, right straight and right curled. Then, by switching different sleeping postures in the six sleeping postures, the number of turning over times is counted. Finally, the above-mentioned monitored physiological signals are combined to comprehensively evaluate the sleep quality. The present invention provides a highly comfortable, multi-parameter fusion, and highly automated sleep monitoring system for the tester, which realizes non-contact monitoring of physiological parameters such as heart rate, respiratory rate, eye movement rate, sleeping posture, and number of turning over, thereby improving the reliability of monitoring. It plays a key role in the clinical diagnosis of sleep quality and the clinical treatment and early intervention of patients or potential patients with sleep disorders.

BRIEF DESCRIPTION OF THE DRAWINGS

Figure 1 is a deep convolutional neural network diagram for physiological signal extraction and amplification

Figure 2 is a heart rate monitoring flow chart

Figure 3 is the eye movement monitoring flow chart

Figure 4 is a flowchart of monitoring sleeping posture and turning over times

Figure 5 is a diagram of the neural network structure for sleeping posture monitoring

Detailed ways

The following is a detailed description of a non-contact sleep monitoring system based on deep learning and multi-parameter fusion according to the present invention in conjunction with the accompanying drawings:

Step 1: Build a sleep monitoring platform. Place three cameras on the upper, left, and right sides of the tester's body to obtain video images of the tester during sleep;

Step 2: Segment the video image obtained by the camera located above the tester's body in step 1 to obtain a video image of the tester's forehead area and a video image of the eye area;

Step 2.1: For the video image captured by the camera above the tester's body, call the dlib library in Python to segment the face area and extract the tester's face video image;

Step 2.2: Perform facial key point detection on the facial video image obtained in step 2.1, call the dlib library in python to perform facial key point detection, and obtain the positions of 68 key points of the tester's face;

Step 2.3: Use the facial key point positions obtained in step 2.2 to segment the video image obtained in step 2.1 again. By identifying the key points of the center of the left and right eyebrows of the face and the upper boundary of the face identified by the dlib library, a rectangular forehead area video image is segmented;

Step 2.4: By identifying the key points of the left and right eyes of the face, find the key points representing the left eye corner, the right eye corner, the uppermost part of the eye socket and the lowermost part of the eye socket, and segment the eye area video image into a rectangular area through the above four key points.

Step 3: Build a deep convolutional neural network for physiological signal extraction and amplification, and use it to extract and amplify tiny physiological signals in the video;

Step 3.1: Build the encoder structure of the deep convolutional neural network for physiological signal extraction and amplification, which includes 2 convolutional layers and 3 residual networks. The Relu activation function is used after each convolutional layer. Then, a convolutional layer is used to extract the physiological signals in the sleep video. The step size is set to 2, and finally the two residual structure outputs are connected;

Step 3.2: Build a modulation amplification structure of a deep convolutional neural network. First, perform a convolution operation on the difference of the physiological signals of the two frames of sleep images through a convolution layer. The activation function is the Relu function, and then multiply it by the amplification factor α. Then use the convolution layer and residual structure to perform nonlinear changes on the amplified features to obtain the amplified physiological signal differential features.

Step 3.3: Build the decoder structure of the deep convolutional neural network, superimpose the amplified physiological signal differential features on the initial sleep image, and then decode and output the amplified video through upsampling and two convolutional layers to achieve the amplification of physiological signals in the sleep video.

Step 4: Input the forehead area video image and the eye area video image obtained in step 2 into the deep convolutional neural network constructed in step 3, respectively, extract and amplify the heart rate signal of the forehead area and the eye movement signal of the eye area, and output the forehead area video image after amplifying the heart rate signal and the eye area video image after amplifying the eye movement frequency;

Step 4.1: Input the forehead area video image obtained in step 2 into the deep convolutional neural network built in step 3, set the amplification factor α=15, extract and amplify the heart rate in the forehead area, and output the forehead area video image after amplifying the heart rate signal;

Step 4.2: Input the eye area video image obtained in step 2 into the deep convolutional neural network built in step 3, set the amplification factor α=30, extract and amplify the eye movement frequency of the eye area, and output the eye area video image after amplifying the eye movement frequency;

Step 5: Separate the RGB three-channels of each frame of the forehead area video image after the amplified heart rate signal obtained in step 4, calculate the average value of the pixels in the three channels of R, G, and B, and then stack the time series to obtain the pulse wave signal;

Step 6: Perform fast Fourier transform on the pulse wave signal obtained in step 5 to obtain a time series spectrum of the human pulse wave;

Step 7: Perform spectrum analysis on the time series spectrum obtained in step 6, and select the frequency corresponding to the spectrum peak as the heart rate monitoring result;

Step 8: stacking the eye region video images obtained in step 4 after amplifying the eye movement frequency in a time series, and performing a fast Fourier transform to obtain a frequency spectrum of the eye region video images;

Step 9: extracting the frequency corresponding to the spectrum peak of the eye area video image spectrum obtained in step 8 as the monitoring result of the eye movement frequency;

Step 10: Build a deep learning-based sleep posture monitoring neural network structure. If the sleep posture monitoring neural network has not been trained, proceed to step 11. If the sleep posture monitoring neural network has been trained, proceed to step 13;

Step 10.1: Build a neural network structure, including 4 convolutional layers, 3 maximum pooling layers, 1 fully connected layer and 1 classifier;

Step 10.2: To prevent the amount of calculation from being too large, extract the key frame of the video at the current moment every 1 second, and extract one frame of image from the cameras located on the upper, left, and right sides of the tester's body respectively, forming three-channel image data and inputting them into the neural network in step 10.1;

Step 10.3: The three-channel image data in step 10.2 are respectively passed through a convolution layer with a convolution kernel of 10×10, a 2×2 maximum pooling layer, a convolution layer with a convolution kernel of 10×10, a 2×2 maximum pooling layer, a convolution layer with a convolution kernel of 10×10, a 2×2 maximum pooling layer, and a convolution layer with a convolution kernel of 10×10 to extract image features;

Step 10.4: Input the above extracted features into the fully connected layer for six classifications. The classification results correspond to six sleeping positions: supine, prone, left side straight, left side curled up, right side straight and right side curled up.

Step 11: Collect more than 1,000 images of the test subject while sleeping by deploying three cameras on the left and right sides of the body, annotate the images with features, and manually annotate the sleeping positions of the test subject, corresponding to six sleeping positions: supine, prone, left side straight, left side curled up, right side straight, and right side curled up;

Step 12: Send the image data annotated in step 11 to the neural network for training, and divide the training set and the validation set into a ratio of 8:2 to train the network until the accuracy of the validation set reaches more than 95%, and the network training is completed;

Step 13: Input the three-dimensional body video images of the tester from the three cameras in step 1 into the trained neural network, and perform six-category output on them, and the classification results correspond to six sleeping positions: supine, prone, left side straight, left side curled up, right side straight, and right side curled up;

Step 14: If the test subject switches between any two of the six sleeping positions in step 13, it is counted as one turn over, but the switch between lying upright on the left side and lying curled up on the left side and the switch between lying upright on the right side and lying curled up on the right side are not counted in the statistics of the number of turns over;

Step 15: Combine the heart rate parameters obtained in step 7, the eye movement frequency obtained in step 9, the sleeping posture obtained in step 13, and the number of tossing and turning obtained in step 14 to comprehensively evaluate the sleep quality results of the tester.

Step 15.1: The heart rate of a person during normal sleep is 60 to 100 beats per minute, and can be reduced to 50 beats per minute during deep sleep. The present invention monitors the heart rate of the test subject during sleep. When the heart rate drops significantly, it is considered that the test subject has entered a deep sleep period, and when the heart rate gradually increases, it is considered that the test subject has exited a deep sleep period. Finally, the proportion of the deep sleep period in the test subject's sleep is counted. The greater the proportion of the deep sleep period, the higher the sleep quality;

Step 15.2: When in the rapid eye movement stage of sleep, the eyeballs will move rapidly. The present invention monitors the eye movements of the test subject. If the eye movement frequency increases significantly, combined with the increase in heart rate in step 14.1, it can be considered that the test subject has entered the rapid eye movement stage. The duration of the rapid eye movement stage is counted. If the rapid eye movement sleep is suddenly interrupted, it is often a signal of the onset of diseases such as angina pectoris and asthma.

Step 15.3: Lying flat is considered to be a better sleeping position during sleep, but it is not suitable for people with respiratory diseases or people who snore frequently. Instead, they should sleep sideways. The present invention monitors the sleeping position of the tester. If the tester suffers from respiratory diseases or snores, the present invention provides sleeping position adjustment suggestions when they adopt a non-side sleeping position.

Step 15.4: The present invention monitors the number of times the tester turns over. If the number of times the tester turns over is too much, it indicates that the tester may be calcium ion deficient or mentally stressed, and has poor sleep quality.

Claims (6)

1. A non-contact sleep monitoring method based on deep learning and multi-parameter fusion, the method comprising the following steps: Step 1: Build a sleep monitoring platform; place three cameras on the upper, left, and right sides of the tester's body to obtain video images of the tester during sleep; Step 2: Segment the video image obtained by the camera located above the tester's body in step 1 to obtain a video image of the tester's forehead area and a video image of the eye area; Step 3: Build a deep convolutional neural network for physiological signal extraction and amplification, and use it to extract and amplify tiny physiological signals in the video; Step 4: Input the forehead area video image and the eye area video image obtained in step 2 into the deep convolutional neural network constructed in step 3, respectively, extract and amplify the heart rate signal of the forehead area and the eye movement signal of the eye area, and output the forehead area video image after amplifying the heart rate signal and the eye area video image after amplifying the eye movement frequency; Step 5: Separate the RGB three-channels of each frame of the forehead area video image after the amplified heart rate signal obtained in step 4, calculate the average value of the pixels in the three channels of R, G, and B, and then stack the time series to obtain the pulse wave signal; Step 6: Perform fast Fourier transform on the pulse wave signal obtained in step 5 to obtain a time series spectrum of the human pulse wave; Step 7: Perform spectrum analysis on the time series spectrum obtained in step 6, and select the frequency corresponding to the spectrum peak as the heart rate monitoring result; Step 8: stacking the eye region video images obtained in step 4 after amplifying the eye movement frequency in a time series, and performing a fast Fourier transform to obtain a frequency spectrum of the eye region video images; Step 9: extracting the frequency corresponding to the spectrum peak of the eye area video image spectrum obtained in step 8 as the monitoring result of the eye movement frequency; Step 10: Build a deep learning-based sleep posture monitoring neural network structure. If the sleep posture monitoring neural network has not been trained, execute step 11; if the sleep posture monitoring neural network has been trained, execute step 13; Step 11: Collect more than 1,000 images of the test subject while sleeping by deploying three cameras on the left and right sides of the body, annotate the images with features, and manually annotate the sleeping positions of the test subject, corresponding to six sleeping positions: supine, prone, left side straight, left side curled up, right side straight, and right side curled up; Step 12: Send the image data annotated in step 11 to the neural network for training, and divide the training set and the validation set into a ratio of 8:2 to train the network until the accuracy of the validation set reaches more than 95%, and the network training is completed; Step 13: Input the three-dimensional body video images of the tester from the three cameras in step 1 into the trained neural network, and perform six-category output on them, and the classification results correspond to six sleeping positions: supine, prone, left side straight, left side curled up, right side straight, and right side curled up; Step 14: If the test subject switches between any two of the six sleeping positions in step 13, it is counted as one turn over, but the switch between lying upright on the left side and lying curled up on the left side and the switch between lying upright on the right side and lying curled up on the right side are not counted in the statistics of the number of turns over; Step 15: Combine the heart rate parameters obtained in step 7, the eye movement frequency obtained in step 9, the sleeping posture obtained in step 13, and the number of tossing and turning obtained in step 14 to comprehensively evaluate the sleep quality results of the tester. 2. A non-contact sleep monitoring system based on deep learning and multi-parameter fusion as claimed in claim 1, characterized in that step 2 specifically comprises: Step 2.1: For the video image captured by the camera above the tester's body, call the dlib library in Python to segment the face area and extract the tester's face video image; Step 2.2: Perform facial key point detection on the facial video image obtained in step 2.1, call the dlib library in python to perform facial key point detection, and obtain the positions of 68 key points of the tester's face; Step 2.3: The video image obtained in step 2.1 is segmented again by using the positions of the key points of the face obtained in step 2.2; a rectangular forehead area video image is segmented by identifying the key points of the center of the left and right eyebrows of the face and the upper boundary of the face identified by the dlib library; Step 2.4: By identifying the key points of the left and right eyes of the face, find the key points representing the left eye corner, the right eye corner, the uppermost part of the eye socket and the lowermost part of the eye socket, and segment the eye area video image into a rectangular area through the above four key points. 3. A non-contact sleep monitoring system based on deep learning and multi-parameter fusion as claimed in claim 1, characterized in that step 3 specifically comprises: Step 3.1: Build the encoder structure of the deep convolutional neural network for physiological signal extraction and amplification, which includes 2 convolutional layers and 3 residual networks. The Relu activation function is used after each convolutional layer. Then, a convolutional layer is used to extract the physiological signals in the sleep video. The step size is set to 2, and finally the two residual structure outputs are connected; Step 3.2: Build a modulation amplification structure of a deep convolutional neural network. First, perform a convolution operation on the difference of the physiological signals of the two frames of sleep images through a convolution layer. The activation function is the Relu function, and then multiply it by the amplification factor α. Then use the convolution layer and residual structure to perform nonlinear changes on the amplified features to obtain the amplified physiological signal differential features. Step 3.3: Build the decoder structure of the deep convolutional neural network, superimpose the amplified physiological signal differential features on the initial sleep image, and then decode and output the amplified video through upsampling and two convolutional layers to achieve the amplification of physiological signals in the sleep video. 4. The non-contact sleep monitoring system based on deep learning and multi-parameter fusion as claimed in claim 1, wherein step 4 specifically comprises: Step 4.1: Input the forehead area video image obtained in step 2 into the deep convolutional neural network built in step 3, set the amplification factor α=15, extract and amplify the heart rate in the forehead area, and output the forehead area video image after amplifying the heart rate signal; Step 4.2: Input the eye area video image obtained in step 2 into the deep convolutional neural network built in step 3, set the amplification factor α=30, extract and amplify the eye movement frequency of the eye area, and output the eye area video image after amplifying the eye movement frequency. 5. The non-contact sleep monitoring system based on deep learning and multi-parameter fusion as claimed in claim 1, wherein step 10 specifically comprises: Step 10.1: Build a neural network structure, including 4 convolutional layers, 3 maximum pooling layers, 1 fully connected layer and 1 classifier; Step 10.2: To prevent the amount of calculation from being too large, extract the key frame of the video at the current moment every 1 second, and extract one frame of image from the cameras located on the upper, left, and right sides of the tester's body respectively, forming three-channel image data and inputting them into the neural network in step 10.1; Step 10.3: The three-channel image data in step 10.2 are respectively passed through a convolution layer with a convolution kernel of 10×10, a 2×2 maximum pooling layer, a convolution layer with a convolution kernel of 10×10, a 2×2 maximum pooling layer, a convolution layer with a convolution kernel of 10×10, a 2×2 maximum pooling layer, and a convolution layer with a convolution kernel of 10×10 to extract image features; Step 10.4: Input the above extracted features into the fully connected layer for six classifications. The classification results correspond to six sleeping positions: supine, prone, left side straight, left side curled up, right side straight and right side curled up. 6. The non-contact sleep monitoring system based on deep learning and multi-parameter fusion as claimed in claim 1, wherein step 15 specifically comprises: Step 15.1: The heart rate of a person during normal sleep is 60 to 100 beats per minute, and can be reduced to 50 beats per minute during deep sleep; the heart rate of the test subject during sleep is monitored, and when the heart rate drops significantly, it is considered that the test subject has entered the deep sleep period, and when the heart rate gradually increases, it is considered that the deep sleep period has been exited; finally, the proportion of the deep sleep period in the test subject's sleep is calculated, and the greater the proportion of the deep sleep period, the higher the sleep quality; Step 15.2: During the REM sleep stage, the eyeballs will move rapidly. Monitor the test subject's eye movements. If the eye movement frequency increases significantly, combined with the heart rate increase in step 14.1, it can be considered that the test subject has entered the REM stage. Count the duration of the REM stage. If the REM sleep is suddenly interrupted, it is often a signal of an attack of diseases such as angina pectoris and asthma. Step 15.3: Lying flat on the back is considered to be a better sleeping position during sleep, but it is not suitable for people with respiratory diseases or frequent snoring. Instead, sleeping on the side should be adopted; monitor the sleeping position of the tester, and if the tester suffers from respiratory diseases or snores, provide sleeping position adjustment suggestions when they adopt a non-side sleeping position; Step 15.4: Monitor the number of times the test subject turns over. If the test subject turns over too often, it indicates that the test subject may be calcium ion deficient or mentally stressed, resulting in poor sleep quality.