Intelligent health old-age care safety monitoring system

Abstract

The invention relates to the technical field of medical or health data information, in particular to an intelligent health old-age care safety monitoring system, which comprises an old-age information acquisition module for acquiring face and basic information; the body health monitoring module predicts health parameters through LSTM, and if the health parameters exceed a threshold value or change abnormally, one is added to the alarm frequency; the psychological health monitoring module uses a StyleGAN to generate an enhanced data set and performs emotion classification through a convolutional neural network, and if the emotion deviates from a baseline, one is added to the alarm frequency; the tumble posture monitoring module judges tumble through posture estimation, and if tumble is detected, one is added to the alarm frequency; the position anomaly monitoring module identifies abnormal movement or residence, and if an anomaly is detected, the number of alarms is increased by one. And guardians and medical staff can obtain alarm information and monitoring reports in time according to the alarm times, so that the efficiency of monitoring the health and safety of the elderly is improved.

Description

Technical Field

[0001] The present invention relates to the field of information technology of medical or health data, and specifically to a smart health and elderly care safety monitoring system. Background Art

[0002] As population aging accelerates, the number of elderly people worldwide is rapidly increasing, placing immense pressure on social security and healthcare systems. Consequently, more and more families are choosing to place their elderly in nursing homes for more specialized care. Nursing homes not only reduce the burden on families but also provide systematic and ongoing health management services for the elderly. However, this care model requires timely and effective monitoring of the health and safety of the elderly without disrupting their daily lives. If changes in their health or emergencies are not addressed promptly, serious consequences can result.

[0003] Therefore, nursing home services urgently need to introduce more intelligent monitoring methods to ensure comprehensive health and safety for seniors. Some existing solutions have achieved some success in health and safety monitoring. For example, smart wearable devices monitor elderly individuals' physiological indicators, such as body temperature, heart rate, and blood pressure, in real time, enabling timely detection of physical abnormalities and improving the efficiency of health management. However, seniors in nursing homes not only face physical deterioration but also often face challenges such as loneliness. If these issues are not adequately addressed, they can lead to exacerbated mental health issues, impacting their overall quality of life. Furthermore, for elderly individuals with unusual movements or special needs, such as falls or prolonged stays, serious physical injuries can occur if not addressed promptly. Most existing solutions still focus primarily on the physical health of seniors, neglecting aspects such as mental health and environmental safety, resulting in a need for improved monitoring efficiency.

[0004] Therefore, a smart health and elderly care safety monitoring system is proposed. Summary of the Invention

[0005] The object of the present invention is to provide a smart and healthy elderly care safety monitoring system. The elderly information acquisition module is used to acquire the facial and basic information of the elderly; the physical health monitoring module collects health parameters, uses the LSTM model to predict future health parameters, and if the real-time and future health parameters exceed the threshold or the change rate is abnormal, an alarm is triggered; the mental health monitoring module separates the expression and skin features through StyleGAN, generates an enhanced data set, and uses a two-dimensional convolutional neural network model to classify the emotions of the enhanced data set and real-time images. If the emotion deviates from the baseline, an alarm is triggered; the fall posture monitoring module judges whether the elderly person has fallen based on posture estimation, and if a fall is detected, an alarm is triggered; the position anomaly monitoring module identifies the abnormal movement or staying state of the elderly person, and if an anomaly is detected, an alarm is triggered. Guardians and medical staff can receive alarm information in a timely manner and receive monitoring reports regularly, improving the monitoring efficiency of the health and safety of the elderly.

[0006] To achieve the above object, the present invention provides the following technical solutions:

[0007] A smart and healthy elderly care safety monitoring system, comprising:

[0008] An elderly information acquisition module, used to acquire the facial information and basic information of the elderly in the nursing home;

[0009] A physical health monitoring module, used to collect health parameters in real time; use the LSTM model to predict future health parameters according to the health parameters; record the health parameters and the future health parameters as the first parameters; if the first parameters exceed the health standard threshold, trigger a first alarm; if the change rate of the first parameters within a preset time window exceeds the standard rate threshold, trigger a second alarm;

[0010] A mental health monitoring module, used to preprocess the facial information to obtain first image data; use the StyleGAN model to separate the facial expression features and skin texture features according to the first image data, and generate an enhanced data set according to the facial expression features; obtain real-time images of the elderly, and use a two-dimensional convolutional neural network model to classify emotions according to the enhanced data set and the real-time images of the elderly to obtain the elderly emotion category; if the probability distribution of the elderly emotion category deviates from the emotion baseline, trigger a third alarm;

[0011] A fall posture monitoring module, used to extract skeleton key points according to the real-time images of the elderly by using a posture estimation method; judge whether the elderly person is in a fall posture according to the skeleton key points, and if the elderly person is in the fall posture, trigger a fourth alarm;

[0012] A position anomaly monitoring module, used to identify any one of abnormal movement and staying of the elderly person.

[0013] Further, the specific process of triggering the first alarm includes:

[0014] The basic information includes age, gender, contact phone number, address, basic medical conditions, and emergency contact information;

[0015] If the elderly person does not have the basic medical condition, classify the elderly person as the first type;

[0016] If the elderly person has the basic medical condition, classify the elderly person as the second type;

[0017] Adjust the health standard threshold according to the first type and the second type, including a first threshold range and a second threshold range; wherein, the second threshold range is smaller than the first threshold range;

[0018] If the first parameter of the first type exceeds the first threshold range, trigger the first alarm;

[0019] If the first parameter of the second type exceeds the second threshold range, trigger the first alarm;

[0020] Set a third threshold range, the third threshold range is smaller than the second threshold range; if the first parameter of the second type does not exceed the second threshold range and exceeds the third threshold range, increase the frequency of collecting the physical health parameters.

[0021] Further, the generation network structure of the StyleGAN model includes:

[0022] An input layer for receiving the first image data;

[0023] A feature separation layer for separating the facial expression feature and the skin texture feature; the feature separation layer includes a low-resolution layer, a medium-resolution layer, and a high-resolution layer; the low-resolution layer is used for separating and extracting the facial expression feature; the medium-resolution layer is used for refining the facial features and expression changes; the high-resolution layer is used for separating and extracting the skin texture feature;

[0024] A generation layer for fixing the skin texture feature and changing the facial expression feature to generate image samples with different facial expressions;

[0025] An output layer for outputting the image samples to obtain the enhanced data set.

[0026] Further, the process of obtaining the real-time elderly image includes:

[0027] Obtain a real-time video stream, and divide the real-time video stream into single-frame images at a second time interval;

[0028] Perform image enhancement on the single-frame image to obtain a first image;

[0029] Use the YOLOv5 model to identify whether the elderly person exists based on the first image. If the elderly person exists, locate the bounding box of the elderly person to obtain the real-time elderly person image.

[0030] Furthermore, the structure of the two-dimensional convolutional neural network model includes:

[0031] An input layer for receiving the enhanced dataset and the real-time elderly person image;

[0032] A convolutional layer for extracting low-level features, intermediate features, and high-level features of the enhanced dataset; the low-level features include the edges and textures of the real-time elderly person image, the intermediate features include the local shapes of the facial features of the real-time elderly person image, and the high-level features include the overall facial expressions of the real-time elderly person image;

[0033] A pooling layer for downsampling the low-level features, the intermediate features, and the high-level features;

[0034] A fully connected layer for extracting the low-level features, the intermediate features, and the high-level features in the pooling layer and performing comprehensive processing to generate a feature vector;

[0035] An output layer for outputting the elderly person's emotion category according to the feature vector.

[0036] Furthermore, the process of establishing the emotion baseline includes:

[0037] Collect the elderly person's emotion category at a first time interval to obtain the probability distribution, which includes a positive emotion distribution and a negative emotion distribution;

[0038] Calculate the change rate of the negative emotion distribution based on the negative emotion distribution at the previous moment and the negative emotion distribution at the current moment;

[0039] If the change rate of the negative emotion distribution exceeds the emotion baseline, trigger the third alarm.

[0040] Furthermore, the fall posture monitoring module specifically includes:

[0041] Based on the real-time elderly person image, use the pose estimation method to extract the skeleton key points, which include the head, shoulders, elbows, knees, and ankles;

[0042] Connect the coordinate points of the head, the coordinate points of the shoulders, the coordinate points of the elbows, the coordinate points of the knees, and the coordinate points of the ankles to obtain an elderly person's skeleton diagram;

[0043] Based on the skeleton diagrams of the elderly person in consecutive frames, calculate the change in the vertical coordinate of the head, the head change speed, and the change in the distance between the shoulders and the knees; if the change in the vertical coordinate of the head exceeds a first threshold, the head change speed exceeds a second threshold, and the distance change exceeds a third threshold, then determine that the elderly person is in a potential falling posture, and input the skeleton diagrams of the elderly person in consecutive frames into a 3D convolutional neural network model to obtain the posture label of the elderly person;

[0044] The posture label includes the falling posture and the non-falling posture; if the elderly person is in the falling posture, trigger the fourth alarm.

[0045] Further, the network structure of the 3D convolutional neural network model includes:

[0046] An input layer for receiving the skeleton diagrams of the elderly person in consecutive frames;

[0047] A convolutional layer for extracting the spatial features and temporal features of the skeleton diagrams of the elderly person; the spatial features are the position changes of the head, the shoulders, the elbows, the knees, and the ankles between consecutive frames; the temporal features include the dynamic processes of the elderly person's body falling, lateral expansion, and lying still flat;

[0048] A pooling layer for downsampling the spatial features and temporal features extracted by the convolutional layer to obtain a dimensionality-reduced feature map;

[0049] A fully connected layer for receiving the dimensionality-reduced feature map processed by the pooling layer and mapping the posture label to the dimensionality-reduced feature map to obtain the predicted score of the posture label;

[0050] An output layer for receiving the predicted score from the fully connected layer.

[0051] Further, the position anomaly monitoring module includes:

[0052] Obtain the position coordinates of the elderly person;

[0053] Set a safe area for the elderly person, calculate the Euclidean distance between the position coordinates of the elderly person and the coordinates of the center point of the safe area, and if the Euclidean distance exceeds the safe radius, trigger the fifth alarm;

[0054] Statistically calculate the residence time of the elderly person based on the position coordinates of the elderly person;

[0055] Further, the system further includes:

[0056] Generate a periodic report according to the physical health monitoring module, the mental health monitoring module, the falling posture monitoring module, and the abnormal position monitoring module, where the periodic report includes the number of alarms and alarm information;

[0057] Divide the elderly into high-risk levels, medium-risk levels, and low-risk levels according to the number of alarms.

[0058] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0059] 1. Regarding the physical health problems of the elderly, by collecting real-time health parameters such as the heart rate and blood pressure of the elderly and using the LSTM model to predict the future health status, abnormal health conditions can be detected in a timely manner to ensure the physical safety of the elderly. In addition, regarding the mental health problems of the elderly, the StyleGAN and two-dimensional convolutional neural network models are used to monitor the mental health status of the elderly through facial expression feature separation and emotion classification. If the emotion significantly deviates from the baseline, an alarm will be triggered. The present invention uses multi-level analysis to help caregivers better pay attention to the psychological changes of the elderly, thereby improving the monitoring efficiency of the health and safety of the elderly.

[0060] 2. The present invention extracts the skeleton key points of the elderly through pose estimation technology to generate a skeleton diagram of the elderly. According to the coordinate changes of the skeleton diagrams of consecutive frames of the elderly, it is judged whether the elderly is in a potential falling state. Once it is detected that the coordinate changes rapidly and presents an abnormal state, the system will input the skeleton diagram into a three-dimensional convolutional neural network model for further analysis to confirm whether it is a falling posture. The present invention can quickly detect falling events and trigger alarms to help the staff handle falling accidents in a timely manner, thereby improving the monitoring efficiency of the health and safety of the elderly.

[0061] 3. By real-time tracking the position coordinates of the elderly, when the elderly exceeds the radius distance of the preset safe area, the system will trigger an alarm to prevent the elderly from getting lost. At the same time, the system will monitor whether the elderly stay at the same location for a long time according to the set residence time threshold. If the residence time threshold is exceeded, the system will trigger an alarm, effectively preventing accidents caused by the elderly staying at the same location for a long time. In addition, the system combines the warning data of each module to generate a periodic monitoring report and divides the risk levels of the elderly to ensure that high-risk elderly are promptly concerned, thereby improving the monitoring efficiency of the health and safety of the elderly. Description of the Drawings

[0062] Figure 1 It is a schematic structural diagram of a smart health elderly care safety monitoring system provided by an embodiment of the present invention;

[0063] Figure 2 It is a schematic diagram of the skeleton of the elderly provided by an embodiment of the present invention.

[0064] In the figure: 1, head; 2, neck; 3, shoulder; 4, elbow; 5, wrist; 6, hip; 7, knee; 8, ankle. Detailed implementation mode

[0065] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0066] Embodiment 1

[0067] In Nursing Home A, as the number of elderly residents continues to increase, the traditional nursing method is gradually becoming inadequate. Although the nursing home has equipped some intelligent hardware, such as intelligent bracelets, security cameras, and health monitoring devices, due to the lack of effective integration of data, the inability to achieve real-time early warning and comprehensive monitoring, the response of nursing staff in case of emergencies is often lagged, and it is difficult to take measures in a timely manner. The physical and mental health of the elderly group is the focus of safety monitoring, and accidental situations such as falls and getting lost are potential risks threatening the safety of the elderly and cannot be ignored. Facing the increasingly complex health needs of the elderly group, Nursing Home A urgently needs a new technical solution to improve the efficiency of health and safety monitoring.

[0068] Please refer to Figures 1 to 2 , the present invention provides a smart health elderly care safety monitoring system, and the technical solution is as follows:

[0069] Figure 1 It is a schematic structural diagram of a smart health elderly care safety monitoring system provided by an embodiment of the present invention.

[0070] As Figure 1 shown, a smart health elderly care safety monitoring system includes:

[0071] An elderly information acquisition module, which is used to acquire the facial information and basic information of the elderly in the nursing home.

[0072] Specifically, when an elderly person is first admitted to the nursing home, with the consent of the elderly person, a dedicated camera device is used to collect the facial image of the elderly person. The nursing staff uses the management terminal to input the basic information of each elderly person, including name, age, gender, contact phone number, address, basic diseases (such as hypertension and diabetes, etc.) and emergency contact information, and stores them in the cloud of the nursing home for reference in subsequent monitoring. In addition, according to the order of admission time of the elderly, a unique number is generated for each elderly person as an identifier of the elderly person's identity. As shown in Table 1, the elderly person's number is bound to the facial information and basic information of the elderly person to facilitate system identification and data management.

[0073] Table 1 Example of Elderly Information

[0074]

[0075] The physical health monitoring module is used to collect health parameters in real time according to the intelligent wearable terminal worn on the elderly; use the LSTM model to predict future health parameters based on the health parameters; record the health parameters and the future health parameters as the first parameters; if the first parameters exceed the health standard threshold, trigger the first alarm; if the change rate of the first parameters within the preset time window exceeds the standard rate threshold, trigger the second alarm.

[0076] Among them, the intelligent wearable terminal can be an intelligent watch, an intelligent bracelet, etc., which can collect various health parameters of the elderly in real time. The health parameters include heart rate, blood pressure, blood oxygen, body temperature, and exercise volume, etc. For data that cannot be collected in real time by the intelligent terminal (such as blood sugar, lung function, etc.), supplementary collection is carried out according to the regular inspections of the nursing staff, and these health parameters are uploaded to the cloud corresponding to the elderly information.

[0077] LSTM (Long Short-Term Memory Network) is a neural network model that can process time series data. Using the LSTM model to predict the health parameters of the elderly can obtain the health status of the elderly in the future period of time.

[0078] Table 2 Comparison of Prediction Model Effects

[0079] model MSE RMSE MAE LSTM model 0.012 0.109 0.075 SVM model 0.025 0.152 0.083 RandomForest model 0.030 0.173 0.110 GRU model 0.015 0.122 0.080

[0080] Specifically, in this embodiment, the hyperparameters of the LSTM model include: the hidden layer size is 128, the number of layers is 3, the batch size is 64, and the learning rate is 0.001. For each elderly person, heart rate, blood pressure, blood oxygen, and body temperature are collected at 15-minute intervals, and blood sugar and lung function are collected at 4-hour intervals. A total of 3 months of various health parameters are collected as historical data, and these historical data are respectively input into the LSTM model for training. The LSTM model will learn the relationship between health parameters and time and predict the health parameters at the next moment. The mean square error (MSE) of the LSTM model is 0.012 and 0.014 respectively, and the training results can meet the requirements.

[0081] To prove the effect of the LSTM model, the SVM model, the Random Forest model, and the GRU model are used as comparison models. As shown in Table 2, the MSE, the root mean square error (RMSE), and the mean absolute error (MAE) all perform the best, indicating that the LSTM model can well predict health parameters.

[0082] Further, the specific process of triggering the first alarm includes:

[0083] The basic information includes age, gender, contact phone number, address, basic medical conditions, and emergency contact information;

[0084] If the elderly person does not have the basic medical conditions (i.e., in good health), then classify the elderly person as the first type;

[0085] If the elderly person has the basic medical conditions (such as diabetes and hypertension, etc.), then classify the elderly person as the second type;

[0086] Adjust the health standard thresholds according to the first type and the second type, including a first threshold range and a second threshold range; wherein, the second threshold range is smaller than the first threshold range;

[0087] If the first parameter of the first type exceeds the first threshold range, then trigger the first alarm;

[0088] If the first parameter of the second type exceeds the second threshold range, then trigger the first alarm;

[0089] Set a third threshold range, the third threshold range is smaller than the second threshold range; if the first parameter of the second type does not exceed the second threshold range but exceeds the third threshold range, then increase the frequency of collecting the physical health parameters.

[0090] Specifically, in this embodiment, for example, for the elderly A (without underlying diseases) and the elderly B (with hypertension), the elderly A is of the first type and the elderly B is of the second type. For the elderly A, the first threshold range is set as follows: heart rate is 60 to 100 beats per minute; blood pressure is 90 / 60 to 140 / 90 mmHg; blood oxygen is 95 to 100%; body temperature is 36.5 - 37.5 °C. If the system monitors that the heart rate of the elderly A is 110 beats per minute, the first alarm will be triggered to remind the caregiver to pay attention to the health status of the elderly. For the elderly B, the second threshold is set as follows: heart rate is 55 to 90 beats per minute; blood pressure is 120 / 80 to 130 / 85 mmHg; blood oxygen is 97 to 100%; body temperature is 36.5 to 37.5 °C. The third threshold is set as follows: heart rate is 60 to 80 beats per minute; blood pressure is 120 / 80 to 130 / 85 mmHg; blood oxygen is 97 to 100%; body temperature is 36.5 to 37 °C. If the system monitors that the heart rate of the elderly B is 88 beats per minute (exceeding the third threshold but not exceeding the second threshold), the time interval will be changed from every 15 minutes to every 5 minutes. If the heart rate of the elderly B returns to normal after 30 minutes of monitoring, the time interval will also be restored to 15 minutes. Here, it is only used as a reference threshold range, and the specific threshold range can be set personalized according to the caregiver's understanding of the health problems of the elderly. Through this hierarchical health management method, it is possible to ensure personalized health monitoring for the elderly with different health conditions, thus improving the monitoring efficiency of the health and safety of the elderly.

[0091] The mental health monitoring module is used to preprocess the facial information to obtain the first image data; separate the facial expression features and skin texture features according to the first image data using the StyleGAN model, and generate an enhanced dataset according to the facial expression features; obtain real-time images of the elderly, and classify the emotions using a two-dimensional convolutional neural network model according to the enhanced dataset and the real-time images of the elderly to obtain the elderly emotion categories; if the probability distribution of the elderly emotion categories deviates from the emotion baseline, trigger the third alarm.

[0092] Among them, the preprocessing process includes:

[0093] Use Gaussian filtering to remove the noise of the facial image to obtain a denoised image; rotate and center the face in the denoised image to ensure that the same face has the same angle and position in different images to obtain a standard image; adjust the brightness and contrast of the standard image to ensure that the facial details are clearly visible, and label each image with tags including seven emotions: anger, disgust, fear, happiness, sadness, surprise, and neutrality, to obtain the first image data.

[0094] The StyleGAN model is an extension of the Generative Adversarial Network (GAN) for generating an enhanced dataset of the elderly. The StyleGAN includes a generator network and a discriminator network, and optimizes the image generation process through adversarial training of the generator network and the discriminator network, specifically including:

[0095] Optimization of the generator network for generating realistic images of the emotions of the elderly, making it impossible for the discriminator network to distinguish whether these images are real;

[0096] Optimization of the discriminator network for distinguishing the differences between real images and generated images of the emotions of the elderly through the cross-entropy loss function.

[0097] Furthermore, the generator network structure of the StyleGAN model includes:

[0098] An input layer for receiving the first image data;

[0099] A feature separation layer for separating the facial expression features and the skin texture features; the feature separation layer includes a low-resolution layer, a medium-resolution layer, and a high-resolution layer. The low-resolution layer is used to separate and extract the facial expression features, such as the overall facial posture and the positions of the eyebrows and eyes; the medium-resolution layer is used to refine the shapes of the facial features and the expression changes, such as a smiling mouth corner and slightly opened eyes, etc.; the high-resolution layer is used to separate and extract the skin texture features, such as wrinkles and skin texture, etc.;

[0100] A generation layer for fixing the skin texture features and changing the facial expression features to generate image samples with different facial expressions;

[0101] An output layer for outputting the image samples to obtain the enhanced dataset.

[0102] Among them, the enhanced dataset includes seven expressions of anger, disgust, fear, happiness, sadness, surprise, and neutral for each elderly person.

[0103] Specifically, traditional emotion recognition usually requires collecting images for each expression one by one, which is difficult in the elderly population. Using the StyleGAN model, seven different expression image samples can be generated based on one or two collected images, enhancing the diversity of the elderly images in the expression dimension. In addition, it also makes the generated expression images more in line with the real facial features of the elderly in terms of details, helping to improve the training effect of the subsequent emotion recognition model, thereby improving the monitoring efficiency of the health and safety of the elderly.

[0104] Furthermore, the process of obtaining the real-time elderly images includes:

[0105] Obtaining a real-time video stream, and splitting the real-time video stream into single-frame images at a second time interval;

[0106] Perform image enhancement on the single-frame image to obtain a first image;

[0107] Use the YOLOv5 model to identify whether the old person exists according to the first image. If the old person exists, locate the bounding box of the old person to obtain the real-time old person image.

[0108] Specifically, obtain a real-time video stream through a camera in the public area of the nursing home. In this embodiment, the second time interval is set to 1 second, which can be customized according to actual needs. Since the overhead of directly processing the video stream is large, the video stream is segmented into single-frame images for frame-by-frame image analysis. Adjust the brightness of the single-frame image, perform Gaussian filtering for denoising, and normalize the resolution to 416×416 pixels. Input the enhanced image into the YOLOv5 model. If a person is detected, output the bounding box to obtain the real-time old person image. Perform face recognition and comparison on the real-time old person image, and use the old person's facial information stored in the old person information acquisition module for identity verification. If the target old person is not matched, it is determined as a staff member or an outsider. In addition, the YOLOv5 model can achieve real-time detection while maintaining high accuracy, which is suitable for the processing requirements of real-time video streams. As shown in Table 3, comparing the YOLOv5 model with the Faster R-CNN model and the SSD model, it can be seen that the YOLOv5 model has higher accuracy and is suitable for detecting whether the old person is in the video. By quickly detecting whether the registered old person in the nursing home exists in the video stream after segmenting it into single frames by YOLOv5, the monitoring efficiency of the health and safety of the old person is improved.

[0109] Table 3 Comparison of the effects of target monitoring models

[0110] model YOLOv5 model FasterR-CNN model SSD model Detection accuracy (mAP@0.5) 90.3％ 88.7％ 84.5％

[0111] Furthermore, the structure of the two-dimensional convolutional neural network model includes:

[0112] An input layer for receiving the enhanced dataset and the real-time old person image;

[0113] A convolutional layer for extracting low-level features, intermediate features, and high-level features of the enhanced dataset; the low-level features include the edges and textures of the real-time old person image, the intermediate features include the local shapes of the facial features of the real-time old person image, and the high-level features include the overall facial expressions of the real-time old person image;

[0114] A pooling layer for downsampling the low-level features, the intermediate features, and the high-level features, capturing the most significant features in the image, and reducing the computational amount;

[0115] A fully connected layer, which is used to extract the low-level features, the mid-level features, and the high-level features in the pooling layer and perform comprehensive processing to generate a feature vector;

[0116] An output layer, which is used to output the elderly emotion category according to the feature vector.

[0117] Specifically, the augmented dataset is divided into 80% and 20% for training and testing the two-dimensional convolutional neural network model. During the training process, the Adam optimizer is used to update the network weights by minimizing the cross-entropy loss function. A total of 100 rounds of training are performed, and the batch size is set to 32. The accuracy rate reaches 92.54%, and it can well identify the emotions in different elderly images. To prove the effect of the two-dimensional convolutional neural network model, the same augmented dataset is used for comparison. As shown in Table 4, the two-dimensional convolutional neural network model also has the best effect during testing. It can efficiently extract the multi-level features of the image and perform accurate classification, thereby improving the monitoring efficiency of the health and safety of the elderly.

[0118] Table 4 Comparison of the effects of prediction models

[0119] model Training accuracy Testing accuracy Two-dimensional convolutional neural network model 92.54％ 91.25％ VGG16 model 91.73％ 90.23％ ResNet-18 model 88.63％ 85.35％

[0120] Furthermore, the process of establishing the emotion baseline includes:

[0121] Collect the elderly emotion category at the first time interval to obtain the probability distribution, which includes the positive emotion distribution and the negative emotion distribution;

[0122] Among them, there are a total of seven expressions: anger, disgust, fear, happiness, sadness, surprise, and neutral. Anger, disgust, fear, and sadness are classified as negative emotions, and happiness, surprise, and neutral are classified as positive emotions; the probability distribution is the ratio of the current emotion category to all emotion categories;

[0123] According to the negative emotion distribution at the previous moment and the negative emotion distribution at the current moment, calculate the negative emotion distribution change rate, which is expressed as:

[0124]

[0125] where, ΔP NEG is the negative emotion distribution change rate, is the value of the i-th negative emotion distribution at the t-th moment, is the value of the i-th negative emotion distribution at the (t - 1)-th moment.

[0126] If the negative emotion distribution change rate exceeds the emotion baseline, trigger the third alarm.

[0127] Specifically, in this embodiment, the first time interval is 24 hours and the emotional baseline is set to 2. Assume that the negative emotion distribution of the elderly person A at the (t - 1)th moment is anger 0.02, sadness 0.05, and anxiety 0.08, and the negative emotion distribution at the tth moment is anger 0.2, sadness 0.5, and anxiety 0.2. Then the change rate of the negative emotion distribution is 5. The change rate of the negative emotion distribution of the elderly person A exceeds the emotional baseline, triggering the third alarm. By continuously tracking and comparing the changes in the negative emotion distribution of the elderly person, the changing trend of their emotions can be monitored in real time. Especially when identifying situations with large emotional fluctuations, the system can issue an alarm in a timely manner, which can help caregivers more accurately understand the emotional state of each elderly person, thereby improving the monitoring efficiency of the health and safety of the elderly.

[0128] The fall posture monitoring module is used to extract the skeleton key points according to the real-time elderly person image using the posture estimation method; determine whether the elderly person is in a fall posture according to the skeleton key points. If the elderly person is in the fall posture, trigger the fourth alarm.

[0129] Specifically, when the elderly person falls, their physical condition may make it difficult for them to call for help on their own. To address the safety issue of possible falls of the elderly in the nursing home, a fall posture monitoring module is introduced.

[0130] Furthermore, the fall posture monitoring module specifically includes:

[0131] Extract the skeleton key points according to the real-time elderly person image using the posture estimation method. The skeleton key points include the head, shoulders, elbows, knees, and ankles;

[0132] Connect the coordinate points of the head, the coordinate points of the shoulders, the coordinate points of the elbows, the coordinate points of the knees, and the coordinate points of the ankles to obtain the elderly person's skeleton diagram;

[0133] According to the elderly person's skeleton diagrams of consecutive frames, calculate the change in the vertical coordinate of the head, the change speed of the head, and the change in the distance between the shoulders and the knees; if the change in the vertical coordinate of the head exceeds the first threshold, the change speed of the head exceeds the second threshold, and the change in the distance exceeds the third threshold, then determine that the elderly person is in a potential fall posture, and input the elderly person's skeleton diagrams of consecutive frames into the 3D convolutional neural network model to obtain the posture label of the elderly person;

[0134] Among them, the change in the vertical coordinate of the head is expressed as:

[0135] ΔY head (t) = Y head (t) - Y head (t - 1);

[0136] Among them, ΔY head(t) is the vertical coordinate change of the head at time t, Y head (t) is the vertical coordinate value of the head at time t, Y head (t - 1) is the vertical coordinate value of the head at time t - 1;

[0137] The head change speed is expressed as:

[0138]

[0139] where A head is the head change speed, ΔY head (t) is the vertical coordinate change of the head at time t, ΔY head (t - 1) is the vertical coordinate change of the head at time t - 1, and Δt is the time difference between time t and time t - 1;

[0140] The distance change between the shoulder and the knee is expressed as:

[0141] D shoudler-knee (t) = ||P shoudler (t) - P knee (t)||;

[0142] ΔD shoudler-knee (t) = D shoudler-knee (t) - D shoudler-knee (t - 1);

[0143] where D shoudler-knee (t) is the distance between the shoulder and the knee at time t, P shoudler (t) is the shoulder coordinate at time t, P knee (t) is the knee coordinate at time t, ||P shoudler (t) - P knee (t)|| is the Euclidean distance at time t, and ΔD shoudler-knee (t) is the distance change between the shoulder and the knee at time t, D shoudler-knee (t - 1) is the distance between the shoulder and the knee at time t - 1.

[0144] The posture label includes the falling posture and the non - falling posture; if the old person is in the falling posture, the fourth alarm is triggered.

[0145] Figure 2 This is a schematic diagram of the skeleton of the old person provided by the embodiment of the present invention.

[0146] As Figure 2As shown, it presents the complete skeletal key points of the elderly. The skeletal key points include the head 1, neck 2, shoulders 3, elbows 4, wrists 5, hips 6, knees 7, and ankles 8. These key points cover the main joint parts of the human body. However, since the goal of the present invention is to focus on the monitoring of the elderly's falling postures, it is not necessary to comprehensively track and analyze all key points. Only the key points closely related to falling are extracted, including the head 1, shoulders 3, elbows 4, knees 7, and ankles 8.

[0147] In this embodiment, the first threshold can be set to 30 pixels, indicating that the head 1 has a relatively obvious downward movement. The second threshold is set to 50 pixels / second, indicating that the head 1 has a relatively fast accelerating downward fall. The third threshold can be set to 20 pixels / second, indicating a significant shortening of the distance between the shoulders 3 and the knees 7. The specific values can be changed according to the actual situation.

[0148] Specifically, the pose estimation method extracts the human skeletal key points, intuitively reflecting the human motion state (such as standing, bending, or falling, etc.). By simplifying the human shape into a skeletal key point model, the data processing volume is reduced. The skeletal key points correspond to the important parts of the human body, can clearly represent the overall pose, and are represented by two-dimensional coordinates. For example, the coordinates of the shoulder 3 are (320, 220), while the coordinates of the knee 7 are (310, 480). When the elderly are in a normal state, these key points usually remain relatively stable; but in abnormal situations such as falling, the coordinates of the head 1 and the distance between the shoulders 3 and the knees 7 will change significantly and rapidly. Combining the analysis of consecutive frames can accurately capture the dynamic process of the falling action, thereby improving the monitoring efficiency of the health and safety of the elderly.

[0149] To prove the effectiveness of the present invention, compared with the two-dimensional convolutional neural network monitoring scheme, the same video stream data is analyzed. The false alarm rate of the two-dimensional convolutional neural network model is 18.8%, while the false alarm rate of the present invention is 13.2%, effectively reducing misjudgment problems such as lying postures and jumping.

[0150] Furthermore, the network structure of the three-dimensional convolutional neural network model includes:

[0151] An input layer for receiving the skeletal diagrams of the elderly in consecutive frames;

[0152] A convolutional layer for extracting the spatial features and temporal features of the skeletal diagrams of the elderly; the spatial features are the position changes of the head 1, the shoulders 3, the elbows 4, the knees 7, and the ankles 8 between consecutive frames; the temporal features include the dynamic processes of the elderly's body falling, lateral expansion, and lying still flat.

[0153] Among them, the body falling is the process of the elderly person changing from an upright posture to a rapid downward movement of the center of gravity, such as the obvious downward movement of the head 1 within 1 second to 2 seconds; the lateral expansion is the lateral spreading of the body after a fall, such as the increase in the lateral distance between the head 1 and the wrist 5; the static lying flat is that the body basically stops moving after a fall, such as the vertical coordinates of the key points of the skeleton being basically at the same level.

[0154] A pooling layer, which is used to downsample the spatial features and the temporal features extracted by the convolutional layer to obtain a feature map with reduced dimensions.

[0155] A fully connected layer, which is used to receive the feature map with reduced dimensions processed by the pooling layer, and map the pose label and the feature map with reduced dimensions to obtain the predicted score of the pose label.

[0156] An output layer, which is used to receive the predicted score of the fully connected layer.

[0157] Specifically, by analyzing the frame sequence, the three-dimensional convolutional neural network model can learn the whole process of the elderly person falling from a standing position to a fallen state. Using the three-dimensional convolutional neural network model can exclude some abnormal but non-falling actions (such as quickly sitting down and bending down to pick up objects), reduce the possibility of misjudgment, improve the reliability of fall recognition, and thus improve the monitoring efficiency of the health and safety of the elderly.

[0158] A position anomaly monitoring module, which is used to identify whether the elderly person has any abnormal movement or residence.

[0159] Furthermore, the position anomaly monitoring module includes:

[0160] Obtain the position coordinates of the elderly person;

[0161] Set a safety area for the elderly person, calculate the Euclidean distance between the position coordinates of the elderly person and the coordinates of the center point of the safety area, and if the Euclidean distance exceeds the safety radius, trigger the fifth alarm;

[0162] Statistical the residence time of the elderly person according to the position coordinates of the elderly person;

[0163] Set a residence threshold for the elderly person, and if the residence time of the elderly person exceeds the residence threshold, trigger the sixth alarm.

[0164] Specifically, some elderly people in nursing homes may be prone to getting lost or unable to judge safe areas due to illnesses (such as Alzheimer's disease and cognitive impairment). In this embodiment, the GPS position coordinates of the elderly are obtained by the intelligent wearable terminal and stored in the monitoring system every 10 seconds. Taking the elderly person's room as the center point of the safe area and the safe radius as 100 meters, that is, a circular area represents the safe range, or the safe radius can be dynamically set according to the areas where the elderly often go. If the elderly person's position exceeds the safe radius, such as walking out of the activity area of the nursing home, an alarm is triggered to notify the guardians or staff for timely handling. In addition, the time that the elderly person stays continuously at a certain point is counted. To avoid long-term stay, the stay time threshold is set at 120 minutes. For example, if an elderly person's position continuously stays in the toilet for more than 2 hours, it may be necessary to check whether their physical condition is abnormal. Through the position anomaly monitoring module, the risk of getting lost or abnormal behavior can be discovered in a timely manner, thereby improving the monitoring efficiency of the health and safety of the elderly.

[0165] Table 5 Example of Periodic Report

[0166]

[0167] Furthermore, the system further includes:

[0168] Generate a periodic report according to the physical health monitoring module, the mental health monitoring module, the fall posture monitoring module, and the position anomaly monitoring module; the periodic report includes the number of alarms and alarm information;

[0169] Classify the elderly into high-risk level, medium-risk level, and low-risk level according to the number of alarms.

[0170] In this embodiment, the periodic report is shown in Table 5, including the number of alarms within a week and the detailed description provided by the alarm handlers, and the report is updated weekly.

[0171] If the total number of alarms is greater than 10 times, the elderly are classified into the high-risk level and need urgent attention; if the total number of alarms is 5 to 9 times, the elderly are classified into the medium-risk level and need key attention; if the total number of alarms is less than 5 times, the elderly are classified into the low-risk level, and the existing monitoring plan is maintained. Among them, the classification level can be adjusted according to the actual situation. Through regular summarization, it helps the staff identify high-risk elderly people and take early intervention measures, and can also formulate targeted health management plans for different elderly people, thereby improving the monitoring efficiency of the health and safety of the elderly.

[0172] Through the physical health monitoring module, the system of the present invention can collect the health data of the elderly in real time and predict future health changes. When the health parameters exceed the set thresholds, it can trigger an alarm in a timely manner, ensuring that the elderly can receive timely attention and intervention and avoiding the deterioration of health problems. The mental health monitoring module can accurately judge the emotional changes of the elderly by analyzing their facial expressions, helping the nursing staff to detect the emotional fluctuations of the elderly in a timely manner, and thus taking psychological intervention measures in a timely manner to improve the mental health of the elderly. The fall posture monitoring module can judge the fall posture of the elderly in real time by extracting the key points of the skeleton of the elderly, and can quickly identify and trigger an alarm to ensure that the elderly can receive timely assistance after falling and reduce the further harm to health caused by the fall. The abnormal position monitoring module can identify whether the elderly enter a dangerous area or have abnormal movement in a timely manner by tracking the position of the elderly in real time, thus preventing the elderly from getting lost or encountering other dangers. With the comprehensive multi-level response mechanism, the present invention improves the monitoring efficiency of the health and safety of the elderly, thus helping the nursing staff to provide a comprehensive intelligent health care plan for the elderly group.

[0173] Embodiment 2

[0174] In Nursing Home B, there are a large number of elderly people, and among the elderly living in Nursing Home B, the proportion of elderly people with advanced age and those suffering from chronic diseases or dementia is high. Although the existing security cameras can provide monitoring images, they lack intelligent analysis capabilities and cannot judge in real time whether the elderly have fallen or strayed into a dangerous area. These elderly people need a more intelligent care plan to reduce the care blind spots and improve the response efficiency.

[0175] A smart health care and safety monitoring system for the elderly includes:

[0176] An elderly information acquisition module for acquiring the facial information and basic information of the elderly in the nursing home;

[0177] A physical health monitoring module for collecting health parameters in real time; predicting future health parameters using an LSTM model according to the health parameters; recording the health parameters and the future health parameters as first parameters; if the first parameters exceed the health standard threshold, triggering a first alarm; if the change rate of the first parameters within a preset time window exceeds the standard rate threshold, triggering a second alarm;

[0178] A mental health monitoring module for preprocessing the facial information to obtain first image data; separating facial expression features and skin texture features using a StyleGAN model according to the first image data, and generating an enhanced data set according to the facial expression features; acquiring a real-time image of the elderly, and performing emotion classification using a two-dimensional convolutional neural network model according to the enhanced data set and the real-time image of the elderly to obtain the elderly emotion category; if the probability distribution of the elderly emotion category deviates from the emotion baseline, trigger a third alarm;

[0179] A fall posture monitoring module for extracting skeleton key points using a posture estimation method according to the real-time image of the elderly; judging whether the elderly is in a fall posture according to the skeleton key points, and if the elderly is in the fall posture, trigger a fourth alarm;

[0180] A location anomaly monitoring module for identifying whether the elderly has any abnormal movement or stay.

[0181] Further, the specific process of triggering the first alarm includes:

[0182] The basic information includes age, gender, contact phone number, address, basic medical conditions and emergency contact information;

[0183] If the elderly does not have the basic medical condition, classify the elderly as the first type;

[0184] If the elderly has the basic medical condition, classify the elderly as the second type;

[0185] Adjust the health standard threshold according to the first type and the second type, including a first threshold range and a second threshold range; wherein, the second threshold range is smaller than the first threshold range;

[0186] If the first parameter of the first type exceeds the first threshold range, trigger the first alarm;

[0187] If the first parameter of the second type exceeds the second threshold range, trigger the first alarm;

[0188] Set a third threshold range, the third threshold range is smaller than the second threshold range; if the first parameter of the second type does not exceed the second threshold range and exceeds the third threshold range, increase the frequency of collecting the physical health parameters.

[0189] Further, the generation network structure of the StyleGAN model includes:

[0190] An input layer for receiving the first image data;

[0191] A feature separation layer for separating the facial expression features and the skin texture features; the feature separation layer includes a low-resolution layer, a medium-resolution layer, and a high-resolution layer; the low-resolution layer is used for separating and extracting the facial expression features; the medium-resolution layer is used for refining the shapes of facial features and expression changes; the high-resolution layer is used for separating and extracting the skin texture features;

[0192] A generation layer for fixing the skin texture features and changing the facial expression features to generate image samples with different facial expressions;

[0193] An output layer for outputting the image samples to obtain the enhanced data set.

[0194] Further, the process of obtaining the real-time elderly image includes:

[0195] Obtaining a real-time video stream, and splitting the real-time video stream into single-frame images at a second time interval;

[0196] Performing image enhancement on the single-frame images to obtain a first image;

[0197] Using the YOLOv5 model to identify whether the elderly person exists in the first image. If the elderly person exists, the bounding box of the elderly person is located to obtain the real-time elderly image.

[0198] Further, the structure of the two-dimensional convolutional neural network model includes:

[0199] An input layer for receiving the enhanced data set and the real-time elderly image;

[0200] A convolutional layer for extracting low-level features, intermediate features, and high-level features of the enhanced data set; the low-level features include the edges and textures of the real-time elderly image, the intermediate features include the local shapes of the facial features of the real-time elderly image, and the high-level features include the overall facial expressions of the real-time elderly image;

[0201] A pooling layer for downsampling the low-level features, the intermediate features, and the high-level features;

[0202] A fully connected layer for extracting the low-level features, the intermediate features, and the high-level features in the pooling layer and performing comprehensive processing to generate feature vectors;

[0203] An output layer for outputting the elderly emotion category according to the feature vectors.

[0204] Further, the process of establishing the emotion baseline includes:

[0205] Collect the emotional categories of the elderly at the first time interval to obtain the probability distribution, which includes the positive emotion distribution and the negative emotion distribution;

[0206] Calculate the change rate of the negative emotion distribution based on the negative emotion distribution at the previous moment and the negative emotion distribution at the current moment;

[0207] If the change rate of the negative emotion distribution exceeds the emotion baseline, trigger the third alarm.

[0208] Furthermore, the fall posture monitoring module specifically includes:

[0209] Extract the skeleton key points using the pose estimation method according to the real-time image of the elderly, and the skeleton key points include the head, shoulders, elbows, knees, and ankles;

[0210] Connect the coordinate points of the head, the coordinate points of the shoulders, the coordinate points of the elbows, the coordinate points of the knees, and the coordinate points of the ankles to obtain the skeleton diagram of the elderly;

[0211] Calculate the change in the vertical coordinate of the head, the change speed of the head, and the change in the distance between the shoulders and the knees based on the skeleton diagrams of the elderly in consecutive frames; if the change in the vertical coordinate of the head exceeds the first threshold, the change speed of the head exceeds the second threshold, and the change in the distance exceeds the third threshold, then determine that the elderly is in a potential fall posture, and input the skeleton diagrams of the elderly in consecutive frames into the 3D convolutional neural network model to obtain the pose label of the elderly;

[0212] The pose label includes the fall posture and the non-fall posture; if the elderly is in the fall posture, trigger the fourth alarm.

[0213] Furthermore, the network structure of the 3D convolutional neural network model includes:

[0214] An input layer for receiving the skeleton diagrams of the elderly in consecutive frames;

[0215] A convolutional layer for extracting the spatial features and temporal features of the skeleton diagrams of the elderly; the spatial features are the position changes of the head, the shoulders, the elbows, the knees, and the ankles between consecutive frames; the temporal features include the dynamic processes of the body falling, lateral expansion, and static lying flat of the elderly;

[0216] A pooling layer for downsampling the spatial features and temporal features extracted by the convolutional layer to obtain a reduced-dimensional feature map;

[0217] The fully connected layer is used to receive the dimensionality-reduced feature map processed by the pooling layer, and map the pose label and the dimensionality-reduced feature map to obtain the predicted score of the pose label;

[0218] The output layer is used to receive the predicted score of the fully connected layer.

[0219] Further, the position anomaly monitoring module includes:

[0220] Obtain the position coordinates of the elderly;

[0221] Set a safety area for the elderly, calculate the Euclidean distance between the position coordinates of the elderly and the coordinates of the center point of the safety area. If the Euclidean distance exceeds the safety radius, trigger the fifth alarm;

[0222] Statistically calculate the residence time of the elderly according to the position coordinates of the elderly;

[0223] Further, the system further includes:

[0224] Generate a periodic report according to the physical health monitoring module, the mental health monitoring module, the fall pose monitoring module and the position anomaly monitoring module; the periodic report includes the number of alarms and alarm information;

[0225] Classify the elderly into high-risk level, medium-risk level and low-risk level according to the number of alarms.

[0226] Specifically, for Nursing Home B, due to the large number of elderly people, in order to reduce the risk of infection, an environmental safety monitoring module can also be added to regularly monitor the disinfection effect in the nursing home.

[0227] The average response time of the alarm in the original monitoring plan of the nursing home is about 5 to 15 minutes. After using the present invention for 3 months, the system analyzes the behavior of the elderly in real time, and notifies the nursing staff by automatically triggering an alarm, and the response time is shortened to 1 to 3 minutes, improving the monitoring efficiency of the health and safety of the elderly, and also providing a strong guarantee for the safety of the elderly.

[0228] Although the embodiments of the present invention have been shown and described, for those of ordinary skill in the art, it can be understood that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principle and spirit of the present invention. The scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A smart and healthy elderly care safety monitoring system, characterized in that, Including: An elderly information acquisition module for acquiring the facial information and basic information of the elderly in the nursing home; A physical health monitoring module for collecting health parameters in real time; Using an LSTM model to predict future health parameters based on the health parameters; recording the health parameters and the future health parameters as first parameters; if the first parameters exceed the health standard threshold, triggering a first alarm; If the change rate of the first parameters within a preset time window exceeds the standard rate threshold, triggering a second alarm; A mental health monitoring module for preprocessing the facial information to obtain first image data; separating the facial expression features and skin texture features according to the first image data using a StyleGAN model, and generating an enhanced data set according to the facial expression features; Obtaining a real-time elderly image, and classifying the emotions of the elderly according to the enhanced data set and the real-time elderly image using a two-dimensional convolutional neural network model to obtain the elderly emotion category; If the probability distribution of the elderly emotion category deviates from the emotion baseline, triggering a third alarm; A fall posture monitoring module for extracting skeleton key points using a pose estimation method according to the real-time elderly image; Judging whether the elderly is in a fall posture according to the skeleton key points, and if the elderly is in the fall posture, triggering a fourth alarm; A position anomaly monitoring module for identifying any one of abnormal movement and staying of the elderly.

2. The intelligent and healthy elderly care safety monitoring system according to claim 1, wherein, The specific process of triggering the first alarm includes: The basic information includes age, gender, contact phone number, address, basic diseases and emergency contact information; If the elderly does not have the basic diseases, classifying the elderly as the first type; If the elderly has the basic diseases, classifying the elderly as the second type; Adjusting the health standard threshold according to the first type and the second type, including a first threshold range and a second threshold range; wherein, the second threshold is smaller than the first threshold range; If the first parameters of the first type exceed the first threshold range, triggering the first alarm; If the first parameters of the second type exceed the second threshold range, triggering the first alarm; Setting a third threshold range, the third threshold range being smaller than the second threshold range; if the first parameters of the second type do not exceed the second threshold range but exceed the third threshold range, increasing the frequency of collecting the physical health parameters.

3. A smart and healthy elderly care safety monitoring system according to claim 1, characterized in that, The generation network structure of the StyleGAN model includes: An input layer for receiving the first image data; A feature separation layer for separating the facial expression features and the skin texture features; the feature separation layer includes a low-resolution layer, a medium-resolution layer and a high-resolution layer; the low-resolution layer is used for separating and extracting the facial expression features; the medium-resolution layer is used for refining the facial features and expression changes; the high-resolution layer is used for separating and extracting the skin texture features; A generation layer for fixing the skin texture features and changing the facial expression features to generate image samples with different facial expressions; An output layer for outputting the image sample to obtain the enhanced data set.

4. A smart and healthy elderly care safety monitoring system according to claim 1, characterized in that, The process of acquiring the real-time elderly image includes: Acquiring a real-time video stream and splitting the real-time video stream into single-frame images at a second time interval; Performing image enhancement on the single-frame image to obtain a first image; Using the YOLOv5 model to identify whether the elderly person exists based on the first image. If the elderly person exists, the bounding box of the elderly person is located to obtain the real-time elderly image.

5. A smart and healthy elderly care safety monitoring system according to claim 1, characterized in that, The structure of the two-dimensional convolutional neural network model includes: An input layer for receiving the enhanced data set and the real-time elderly image; A convolutional layer for extracting low-level features, intermediate features, and high-level features of the enhanced data set; the low-level features include the edges and textures of the real-time elderly image, the intermediate features include the local shapes of the facial features of the real-time elderly image, and the high-level features include the overall facial expressions of the real-time elderly image; A pooling layer for downsampling the low-level features, the intermediate features, and the high-level features; A fully connected layer for extracting the low-level features, the intermediate features, and the high-level features in the pooling layer and performing comprehensive processing to generate a feature vector; An output layer for outputting the elderly emotion category according to the feature vector.

6. The intelligent and healthy elderly care safety monitoring system according to claim 1, characterized in that The process of establishing the emotion baseline includes: Collecting the elderly emotion category at a first time interval to obtain the probability distribution, where the probability distribution includes a positive emotion distribution and a negative emotion distribution; Calculating the change rate of the negative emotion distribution based on the negative emotion distribution at the previous moment and the negative emotion distribution at the current moment; If the change rate of the negative emotion distribution exceeds the emotion baseline, the third alarm is triggered.

7. The intelligent and healthy elderly care safety monitoring system according to claim 1, characterized in that, The fall posture monitoring module specifically includes: Extracting the skeleton key points using the pose estimation method based on the real-time elderly image, where the skeleton key points include the head, shoulders, elbows, knees, and ankles; Connecting the coordinate points of the head, the coordinate points of the shoulders, the coordinate points of the elbows, the coordinate points of the knees, and the coordinate points of the ankles to obtain an elderly skeleton diagram; Calculating the change in the vertical coordinate of the head, the change speed of the head, and the change in the distance between the shoulders and the knees based on the elderly skeleton diagrams of consecutive frames; if the change in the vertical coordinate of the head exceeds a first threshold, the change speed of the head exceeds a second threshold, and the change in the distance exceeds a third threshold, it is determined that the elderly person is in a potential fall posture, and the elderly skeleton diagrams of consecutive frames are input into a three-dimensional convolutional neural network model to obtain the pose label of the elderly person; The pose label includes the fall posture and the non-fall posture; if the elderly person is in the fall posture, the fourth alarm is triggered.

8. The intelligent and healthy elderly care safety monitoring system according to claim 7, wherein The network structure of the three-dimensional convolutional neural network model includes: An input layer for receiving the elderly skeleton diagrams of consecutive frames; A convolutional layer for extracting the spatial and temporal features of the skeleton diagram of the elderly; the spatial features are the position changes of the head, shoulders, elbows, knees, and ankles between consecutive frames; the temporal features include the dynamic processes of the elderly's body falling, lateral expansion, and lying still flat. A pooling layer for downsampling the spatial and temporal features extracted by the convolutional layer to obtain a dimensionality-reduced feature map. A fully connected layer for receiving the dimensionality-reduced feature map processed by the pooling layer and mapping the pose label to the dimensionality-reduced feature map to obtain the predicted score of the pose label. An output layer for receiving the predicted score from the fully connected layer.

9. The intelligent and healthy elderly care safety monitoring system according to claim 1, characterized in that The position anomaly monitoring module includes: Obtaining the position coordinates of the elderly. Setting a safe area for the elderly, calculating the Euclidean distance between the position coordinates of the elderly and the coordinates of the center point of the safe area, and triggering a fifth alarm if the Euclidean distance exceeds the safety radius. Statistical the residence time of the elderly according to the position coordinates of the elderly. Setting a residence threshold for the elderly, and triggering a sixth alarm if the residence time of the elderly exceeds the residence threshold.

10. A smart and healthy elderly care safety monitoring system according to claim 1, characterized in that, The system further includes: Generating a periodic report according to the physical health monitoring module, the mental health monitoring module, the fall pose monitoring module, and the position anomaly monitoring module, where the periodic report includes the number of alarms and alarm information. Classifying the elderly into high-risk level, medium-risk level, and low-risk level according to the number of alarms.

Images