Multi-mode unsupervised health monitoring system based on edge cloud collaboration

Abstract

The invention belongs to the technical field of artificial intelligence and Internet of Things, and discloses a multi-mode unsupervised health monitoring system based on edge cloud collaboration, and the specific technical scheme is as follows: the system comprises a data collection layer, and the data collection layer, an edge calculation layer, a cloud analysis layer and an application layer communicate through data interfaces; the data acquisition layer is used for acquiring human body motion images in real time, the data acquisition layer undertakes a front-end data collection task, and the edge calculation layer locates a human body by using a YOLO target detection algorithm, realizes cross-frame individual identity continuous recognition through a DeepSort tracking algorithm, generates motion track information of each monitored person, and transmits the motion track information to the data acquisition layer. The cloud analysis layer serves as a back-end platform and is responsible for executing behavior recognition and health risk modeling and an application layer, the application layer achieves information interaction and early warning linkage functions for terminal users and related managers, when high-risk behaviors are detected, multi-channel linkage alarm is conducted immediately, and meanwhile event detail records are generated on a user platform.

Description

Technical Field

[0001] The present invention belongs to the technical field of artificial intelligence and the Internet of Things, and particularly relates to a multi-modal unsupervised health monitoring system based on edge-cloud collaboration. Background Art

[0002] In recent years, with the increasing demand for health monitoring, anomaly detection systems relying on computer vision and artificial intelligence technologies have received increasing attention. Especially in the context of an aging society, real-time analysis of health-related behaviors such as falls and abnormal gaits of individuals through video surveillance has important social value. However, existing health monitoring systems still face many challenges in practical applications.

[0003] Currently, most health monitoring systems rely on uploading the collected raw video data to a cloud server for centralized processing. Although this method improves the computing efficiency, it also brings a serious risk of privacy leakage. This problem has become more prominent after the introduction of laws such as the Data Protection Regulation, especially in sensitive scenarios such as homes and nursing homes, where the requirements for data privacy are more stringent.

[0004] In addition, anomaly detection systems usually rely on supervised learning methods in deep learning, which requires a large number of labeled anomaly behavior samples to achieve a high recognition accuracy. However, in reality, many abnormal behaviors are low-frequency and highly diverse. For example, gait features related to chronic diseases are difficult to obtain sufficient training data through traditional acquisition methods. This makes the model have obvious deficiencies in dealing with unknown abnormal behaviors and limited generalization ability.

[0005] To extract spatio-temporal features, some current studies propose to use models such as the spatio-temporal graph convolutional network (ST-GCN). As described in the literature "Spatial-Temporal GCN for Action Recognition" (AAAI 2020), this type of model usually introduces a 3D convolutional structure, and its model parameters are as high as 10MB. Although it can effectively model the spatio-temporal relationship between human joint points, it also significantly increases the computational complexity of the model and cannot be deployed in real time on edge devices, limiting its practicality.

[0006] At the same time, current health monitoring systems are mostly single-functional independent modules. For example, the fall detection system proposed in the patent with the patent publication number CN119296172A and the patent name "Fall Detection System Based on AI Image Analysis" has basic anomaly detection functions. However, due to its high dependence on raw video transmission and single-task processing ability, it lacks means for collaborative analysis of multi-modal data and is difficult to achieve a comprehensive understanding and prediction of complex health states.

[0007] In summary, the existing technologies still have significant deficiencies in privacy protection, data dependence, model deployment, and multi-functional integration, and there is an urgent need for innovative breakthroughs in directions such as algorithm design, system integration, and edge computing. Summary of the Invention

[0008] To solve the technical problems existing in the prior art, the present invention provides a health monitoring system and method for nursing homes and home care scenarios. The system collaborates between edge computing and the cloud, integrates unsupervised anomaly detection and chronic disease prediction functions, and realizes multi-modal health monitoring with high privacy and low resource consumption.

[0009] To achieve the above object, the technical solution adopted by the present invention is: a multi-modal unsupervised health monitoring system based on edge-cloud collaboration. The specific system structure is as follows: it includes a data acquisition layer, an edge computing layer, a cloud analysis layer, and an application layer. The data acquisition layer, the edge computing layer, the cloud analysis layer, and the application layer communicate through data interfaces;

[0010] The data acquisition layer is used to collect human activity images in real time. The data acquisition layer undertakes the task of collecting front-end data and has an edge caching mechanism to temporarily store the original video stream for edge processing. The frame rate of the collected images is not less than 15 FPS, and the resolution is not less than 1080P;

[0011] The edge computing layer uses the YOLO object detection algorithm to locate the human body and realizes the continuous identification of individual identities across frames through the DeepSort tracking algorithm, generating the motion trajectory information of each monitored person. The system calls the OpenPose algorithm to extract 17 key skeletal point information of the human body. The key point data is structured and encoded into a lightweight JSON format. All uploaded data is desensitized structured pose information and is encrypted by the AES-256 symmetric encryption algorithm during transmission;

[0012] The cloud analysis layer, as the back-end platform, is responsible for performing behavior recognition and health risk modeling. In terms of abnormal behavior recognition, the system constructs a lightweight spatio-temporal graph convolutional network ST-GCN based on skeletal key points to process the human action sequence in a temporal manner, and introduces a spatio-temporal normalizing flow model STG-NF to model the probability density of skeletal key points;

[0013] In terms of chronic disease analysis, the system constructs a gait feature sequence based on the uploaded pose key point data. Through the combined modeling method of the lightweight spatio-temporal graph convolutional network ST-GCN and the spatio-temporal normalizing flow model STG-NFF, the system shows the topological relationship and time evolution law between joints in individual movements, thereby inferring the early characteristic manifestations of diseases. This module provides auxiliary diagnosis suggestions for medical services or rehabilitation institutions, realizing the dynamic monitoring and trend prediction of individual health status;

[0014] The application layer is oriented to end-users and relevant management personnel, realizing information interaction and early warning linkage functions. When the system detects high-risk behaviors, it immediately conducts multi-channel linkage alarms and generates event detail records on the user platform simultaneously.

[0015] The system generates a posture heat map and behavior logs every day, providing a 24-hour visual display of behavior trends to help users identify the frequency and pattern changes of potential risk behaviors. The results of chronic disease assisted analysis are fed back to medical staff or family members in the form of reports.

[0016] In the data acquisition layer, the system only extracts the two-dimensional coordinate sequence composed of skeleton key points and caches it locally in the form of an anonymized structure. The original video images are not uploaded or stored. All stored and transmitted data adopt the AES-256 encryption standard and a local regular destruction policy is provided.

[0017] In the edge computing layer, the system first calls the lightweight YOLOv5-Nano model to detect human targets in each frame of the image. This module is based on the CSPDarknet feature extraction network, uses the SiLU activation function and the depthwise separable convolution structure, and its output is the bounding box coordinates and classification probabilities of human targets, providing an input basis for subsequent tracking and pose estimation.

[0018] After detection, the system uses the DeepSort algorithm to perform cross-frame matching and identity tracking on human targets. This module integrates a Kalman filter predictor and a ReID embedded feature matcher, assigns an independent ID to each detected target, and maintains its temporal trajectory.

[0019] In the cloud analysis layer, the ST-GCN pose modeling module uses OpenPose technology to extract the human skeleton key points in each frame of the image, forming a temporal graph structure. The ST-GCN model conducts spatio-temporal joint modeling on this skeleton graph to capture the motion states and action features of individuals over a period of time.

[0020] The human skeleton is modeled as a graph G=(V,E) in each frame. Among them, the node V represents the key points, and the edge E represents the joint connection relationship and the temporal adjacency relationship. The propagation form of each layer of graph convolution is:

[0021]

[0022] Among them, f in (v j ) is the input feature, f out (v i ) is the output feature, W l (v i ,v j ) is the convolution weight, Zij is the normalization factor, is the neighbor set of node v i ;

[0023] The system introduces the STG-NF normalizing flow model, which takes the ST graph structure as input, constructs an implicit spatio-temporal action distribution, maps the skeleton data to the standard normal space through a transformation function, and the system evaluates the log-likelihood value of the input action sequence to determine whether it is an abnormal behavior or an early chronic disease posture feature;

[0024] The specific structure of the STG-NF module is as follows:

[0025]

[0026] where p(z) is the probability density of the standard normal distribution, is the Jacobian matrix.

[0027] In the cloud analysis layer, the system is based on the lightweight spatio-temporal graph convolutional network ST-GCN. Channel pruning and sparse graph structure design are introduced in the spatial convolution module, and the YOLOv5-Nano version is selected for the object detection part to streamline the feature channels and the detection head structure. The volume of all models is controlled within 1MB;

[0028] The system adopts an 8-bit integer quantization method in cooperation with the TensorRT inference acceleration module. The specific formula is as follows:

[0029]

[0030] x ≈ s · (x q - z)

[0031] where S is the scaling factor, z is the zero point, round(.) is the rounding function, and x q is the quantized integer value, and it is agreed that x q ∈ [0, 255] or x q ∈ [-128, 127];

[0032] The knowledge distillation mechanism is used to guide the student model to learn the skeleton action feature distribution by the teacher model. The distillation loss function is composed of the KL divergence between the output of the student model and the soft label of the teacher model. The specific formula is as follows:

[0033]

[0034] where y s is the output of the student model, y is the true label, z s , z t are the logits of the student and teacher models respectively, T is the temperature coefficient, and α is the weight coefficient.

[0035] Compared with the prior art, the specific beneficial effects of the present invention are as follows:

[0036] I. Strong privacy compliance: The system does not collect sensitive data that can identify identities throughout the process, complies with international standards such as GDPR, and is applicable to highly sensitive scenarios.

[0037] II. High resource efficiency: The model is highly streamlined and can achieve real-time recognition and processing on resource-constrained edge devices.

[0038] III. High algorithm integration: The system realizes deep coordination of functions such as object detection, object tracking, pose recognition, and risk assessment.

[0039] IV. Strong function integration: The system has multiple functions such as fall detection, abnormal behavior recognition, and chronic disease auxiliary analysis, and has good expansion potential.

[0040] V. Wide range of applicable scenarios: The system can be deployed in various actual scenarios such as homes, hospitals, nursing homes, industrial workshops, mines, etc. The technical solution provided by the present invention has clear innovation, practicability, and popularization, and is of great significance for promoting the development of intelligent health monitoring and behavior recognition technologies. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 is the system architecture diagram of the present invention.

[0042] Figure 2 is the flow chart of the ST-GCN + normalization flow joint model.

[0043] Figure 3 is the schematic diagram of the privacy protection process. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0044] In order to make the technical problems, technical solutions, and beneficial effects to be solved by the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0045] The multi-modal unsupervised health monitoring system based on edge-cloud collaboration aims to realize real-time detection of abnormal behaviors such as falls and convulsions, as well as early risk prediction and management of various chronic diseases on the premise of protecting personal privacy. The system architecture is as Figure 1 shown, and is divided into four levels as a whole, including a data acquisition layer, an edge computing layer, a cloud analysis layer, and an application layer. Communication between each layer is carried out through a structured and encrypted data interface to ensure the unity of processing efficiency and data security.

[0046] Data Acquisition Layer: The system deploys image acquisition devices with high frame rate and high resolution in the target area for real-time acquisition of human activity images. The typical configuration is a frame rate of not less than 15 FPS and a resolution of not less than 1080P to meet the requirements of subsequent pose recognition for image quality. The data acquisition layer not only undertakes the task of front-end data collection but also has an edge caching mechanism to temporarily save the original video stream for edge processing. To prevent privacy leakage, the system is equipped with an image local destruction mechanism, and the original data is erased immediately after processing, without entering the cloud and not being stored on the local disk, realizing the minimization of the retention of the original images.

[0047] Edge Computing Layer: The edge computing layer is mainly responsible for the intelligent preprocessing and privacy protection of the acquired images and is a key intermediate link in the whole system. This layer first uses the YOLO object detection algorithm to accurately locate the human body and realizes the continuous recognition of individual identities across frames through the DeepSort tracking algorithm, generating the motion trajectory information of each monitored person. On this basis, the system calls the OpenPose algorithm to extract the information of 17 key skeletal points of the human body, including the positions of the head, torso, limbs and other parts. To reduce bandwidth occupancy and protect privacy, the key point data is structured and encoded into a lightweight JSON format. The entire edge-side processing process does not retain the original image, and all uploaded data are desensitized structured pose information, which is encrypted by the AES-256 symmetric encryption algorithm during transmission, in line with international common privacy protection regulations.

[0048] Cloud Analysis Layer: The cloud, as a high-performance backend platform, is responsible for executing complex behavior recognition and health risk modeling algorithms. In terms of abnormal behavior recognition, the system constructs a lightweight spatio-temporal graph convolutional network (ST-GCN) based on skeletal key points to process the human action sequence in a temporal manner and achieve high-precision behavior discrimination with a relatively low model complexity. On this basis, as Figure 2 shown, the spatio-temporal normalizing flow model (STG-NF) is introduced. By modeling the probability density of skeletal key points, the system further enhances its recognition ability for low-frequency abnormal behaviors (such as falls and convulsions), and can maintain high robustness and adaptability even in the face of individual differences and environmental interferences.

[0049] In terms of chronic disease analysis, the system constructs a gait feature sequence based on the uploaded pose key point data, including multiple parameters such as stride, stride frequency, steering angle stability, and acceleration change within a period. Through the combined modeling method of ST-GCN and STG-NF, the system explicitly learns the topological relationship and time evolution law between joints in individual movements, so as to infer the early characteristic manifestations of diseases such as Alzheimer's disease, multiple sclerosis, osteoarthritis, amyotrophic lateral sclerosis, etc. This module can provide auxiliary diagnosis suggestions for medical services or rehabilitation institutions, realizing the dynamic monitoring and trend prediction of individual health status.

[0050] Application layer: The application layer is oriented to end-users and relevant management personnel, mainly realizing information interaction and early warning linkage functions. When the system detects high-risk behaviors such as suspected falls and convulsions, it can immediately conduct multi-channel linkage alarms through text messages, voice broadcasts, warning lights, etc., and at the same time generate event details records on the user platform. In addition, the system can generate posture heat maps and behavior logs every day, providing a 24-hour visual display of behavior trends to help users identify the frequency and pattern changes of potential risk behaviors. The results of chronic disease auxiliary analysis can also be fed back to medical staff or family members in the form of reports to achieve early detection and early intervention.

[0051] Lightweight modeling and edge inference optimization: Since this system needs to be deployed on resource-constrained edge devices (such as embedded motherboards, smart cameras), a set of efficient lightweight model construction and inference systems are designed, including key strategies such as model structure pruning, parameter compression, integer quantization, and knowledge distillation to ensure stable operation on low-power platforms.

[0052] First of all, the system is based on the ST-GCN graph neural network. Channel pruning and sparse graph structure design are introduced in the spatial convolution module, significantly reducing the number of parameters and computational overhead. For the object detection part, the YOLOv5-Nano version is selected, and the feature channels and detection head structure are further streamlined. The volume of all models is controlled within 1MB, and the inference speed can reach more than 10 frames per second on platforms such as Raspberry Pi and Jetson Nano.

[0053] At the same time, the system adopts 8-bit integer quantization (INT8) technology, combined with the TensorRT inference acceleration module to improve the deployment efficiency. The formula is as follows:

[0054]

[0055] x≈s·(x q -z)

[0056] where S is the scaling factor; z is the zero point; round(.) is the rounding function; x q is the quantized integer value. For INT8 quantization, it is usually agreed that x q ∈[0,255] or x q ∈[-128,127].

[0057] In addition, through the knowledge distillation mechanism, the high-performance teacher model is used to guide the student model to learn the skeleton action feature distribution, effectively improving the accuracy and stability of the lightweight model. The distillation loss function is composed of the KL divergence between the output of the student model and the soft label of the teacher model, and the form is as follows:

[0058]

[0059] Among them, y s is the output of the student model, y is the true label, and z s and z t are the logits of the student and teacher models respectively, T is the temperature coefficient, and α is the weight coefficient.

[0060] This embodiment realizes the ability to deploy a complete behavior recognition system on edge devices, providing technical support for application scenarios such as home security and industrial safety.

[0061] To meet the requirements of privacy regulations, as Figure 3 shown, the system introduces an end-side data minimization and privacy protection mechanism from the very beginning of its design. Its core concept is: only collect human body posture data, and do not collect any sensitive information that can identify identities such as facial features, background images, and voice information.

[0062] In the data collection stage, the system only extracts the two-dimensional coordinate sequence composed of skeleton key points (such as shoulders, elbows, knees, ankles, etc.) and caches it locally in the form of an anonymized structure; the original video images are not uploaded or stored to ensure irreversible traceability of information. All stored and transmitted data uses the AES-256 encryption standard, and a local regular destruction policy is provided to further protect user privacy.

[0063] This mechanism not only improves the adaptability of the system in highly sensitive scenarios such as medical care, elderly care, and families, but also reduces the data compliance risk, which is one of the important innovation points of the present invention.

[0064] Design of a multi-algorithm collaborative fusion human behavior recognition system. This system adopts a multi-module hierarchical collaborative structure design, integrating multiple key algorithm modules such as object detection, object tracking, pose modeling, and abnormal behavior recognition, forming a complete closed-loop from the original video to behavior judgment. The main process includes the following four sub-modules: YOLO object detection module: The system first calls the lightweight YOLOv5-Nano model to quickly detect human targets in each frame of the image. This module is based on the CSPDarknet feature extraction network, uses the SiLU activation function and the depthwise separable convolution structure, and has high efficiency and high accuracy. Its output is the bounding box coordinates and classification probabilities of human targets, providing the input basis for subsequent tracking and pose estimation.

[0065] DeepSort object tracking module: After detection, the system uses the DeepSort algorithm to perform cross-frame matching and identity tracking on human targets. This module integrates a Kalman filter predictor and a ReID embedding feature matcher, which can assign an independent ID to each detected target and maintain its temporal trajectory. This module ensures the coherent recognition ability of a single individual in the video.

[0066] ST-GCN Pose Modeling Module: With the help of OpenPose technology, the human skeletal key points in each frame of the image are extracted to form a temporal graph structure. The ST-GCN model performs spatio-temporal joint modeling on this skeleton graph to capture the motion state and action features of an individual over a period of time. This module is mainly used to identify basic behavior categories, such as normal walking, standing, falling, etc. Specifically, the human skeleton is modeled as a graph G=(V, E) in each frame, where the nodes V represent the key points and the edges E represent the joint connection relationships and temporal adjacency relationships. The propagation form of each layer of graph convolution is:

[0067]

[0068] where f in (v j ) is the input feature; f out (v i ) is the output feature; W l (v i , v j ) is the convolution weight; Z ij is the normalization factor; is the neighbor set of node v i .

[0069] STG-NF Implicit Behavior Modeling Module: To further identify potential health risk behaviors, the system introduces the STG-NF normalizing flow model. This model takes the ST graph structure as the input, constructs an implicit spatio-temporal action distribution, and maps the skeleton data to the standard normal space through a transformation function. The system evaluates the log-likelihood value of the input action sequence to determine whether it is an abnormal behavior or an early chronic disease posture feature (such as unsteady gait, stiff movement, etc.).

[0070] The STG-NF module has strong abnormal modeling capabilities and can be applied to various application scenarios such as home health monitoring, elderly fall warning, and chronic disease assisted screening.

[0071]

[0072] where p(z) is the probability density of the standard normal distribution;; is the Jacobian matrix.

[0073] This system does not collect sensitive data that can identify identities throughout the process, complies with international standards such as GDPR, and is applicable to highly sensitive scenarios; the model is highly streamlined and can achieve real-time recognition and processing on resource-constrained edge devices; the system realizes deep coordination of functions such as object detection, object tracking, pose recognition, and risk assessment; the system has multiple functions such as fall detection, abnormal behavior recognition, and chronic disease assisted analysis, and has good expansion potential; the system can be deployed in various actual scenarios such as homes, hospitals, nursing homes, industrial workshops, mines, etc. The technical solution provided by the present invention has clear innovation, practicability, and popularization, and is of great significance to promoting the development of intelligent health monitoring and behavior recognition technologies.

[0074] The above are only the preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent replacements, and improvements made within the spirit and principle of the present invention shall be included within the scope of the present invention.

Claims

1. A multi-modal unsupervised health monitoring system based on edge-cloud collaboration, characterized in that The specific system structure is as follows: It includes a data acquisition layer, an edge computing layer, a cloud analysis layer, and an application layer. The data acquisition layer, the edge computing layer, the cloud analysis layer, and the application layer communicate with each other through data interfaces; The data acquisition layer is used to collect human activity images in real time. The data acquisition layer undertakes the task of front-end data collection and has an edge caching mechanism to temporarily save the original video stream for edge processing. The frame rate of the collected images is not lower than 15 FPS, and the resolution is not lower than 1080P; The edge computing layer uses the YOLO target detection algorithm to locate the human body and realizes the continuous identification of individual identities across frames through the DeepSort tracking algorithm, generating the motion trajectory information of each monitored person. The system calls the OpenPose algorithm to extract the information of 17 key skeletal points of the human body. The key point data is structured and encoded into a lightweight JSON format. All uploaded data is desensitized structured pose information and is encrypted by the AES-256 symmetric encryption algorithm during the transmission process; The cloud analysis layer, as the backend platform, is responsible for performing behavior recognition and health risk modeling. In terms of abnormal behavior recognition, the system constructs a lightweight spatio-temporal graph convolutional network ST-GCN based on skeletal key points to process the human action sequence in a temporal manner, and introduces the spatio-temporal normalizing flow model STG-NF to model the probability density of the skeletal key points; In terms of chronic disease analysis, the system constructs a gait feature sequence based on the uploaded pose key point data. Through the joint modeling method of the lightweight spatio-temporal graph convolutional network ST-GCN and the spatio-temporal normalizing flow model STG-NFF, the system shows the topological relationship and time evolution law between joints in individual movements, so as to infer the early characteristic manifestations of diseases. This module provides auxiliary diagnosis suggestions for medical services or rehabilitation institutions to realize the dynamic monitoring and trend prediction of individual health status; The application layer is oriented to end users and relevant management personnel to realize information interaction and early warning linkage functions. When the system detects high-risk behaviors, it immediately conducts multi-channel linkage alarms and generates event detail records on the user platform at the same time; The system generates a pose heat map and a behavior log every day to provide a 24-hour visual display of behavior trends, helping users identify the frequency and pattern changes of potential risk behaviors. The results of chronic disease auxiliary analysis are fed back to medical staff or family members in the form of reports.

2. The multimodal unsupervised health monitoring system based on edge-cloud collaboration according to claim 1, characterized in that, In the data acquisition layer, the system only extracts the two-dimensional coordinate sequence composed of skeletal key points and caches it locally in the form of an anonymized structure. The original video images are not uploaded or stored. All stored and transmitted data uses the AES-256 encryption standard and a local regular destruction strategy is provided.

3. The multimodal unsupervised health monitoring system based on edge-cloud collaboration according to claim 2, wherein In the edge computing layer, the system first calls the lightweight YOLOv5-Nano model to detect human targets in each frame of image. This module is based on the CSPDarknet feature extraction network, uses the SiLU activation function and the depthwise separable convolution structure, and its output is the bounding box coordinates and classification probability of the human target, providing an input basis for subsequent tracking and pose estimation; After the detection is completed, the system uses the DeepSort algorithm to perform cross-frame matching and identity tracking on human targets. This module integrates a Kalman filter predictor and a ReID embedded feature matcher, assigns independent IDs to each detected target, and maintains its temporal trajectory.

4. The multimodal unsupervised health monitoring system based on edge-cloud collaboration according to claim 3, wherein In the cloud analysis layer, the ST-GCN pose modeling module uses OpenPose technology to extract the human skeleton key points in each frame of the image, forming a temporal graph structure. The ST-GCN model performs spatio-temporal joint modeling on this skeleton graph to capture the motion state and action features of an individual over a period of time; The human skeleton is modeled as a graph G=(V, E) in each frame, where the nodes V represent the key points, and the edges E represent the joint connection relationship and the temporal adjacency relationship. The propagation form of each layer of graph convolution is: Among them, f in (v j ) is the input feature, f out (v i ) is the output feature, W l (v i , v j ) is the convolutional weight, Z ij is the normalization factor, is the neighbor set of node v i ; The system introduces the STG-NF normalizing flow model, which takes the ST graph structure as the input, constructs an implicit spatio-temporal action distribution, maps the skeleton data to the standard normal space through a transformation function, and the system evaluates the log-likelihood value of the input action sequence to determine whether it is an abnormal behavior or an early chronic disease pose feature; The specific structure of the STG-NF module is: where p(z) is the probability density of the standard normal distribution, is the Jacobian matrix.

5. The multi-modal unsupervised health monitoring system based on edge-cloud collaboration according to claim 4, wherein, In the cloud analysis layer, the system is based on the lightweight spatio-temporal graph convolutional network ST-GCN. Channel pruning and sparse graph structure design are introduced in the spatial convolution module, and the YOLOv5-Nano version is selected for the object detection part to streamline the feature channels and the detection head structure. The volume of all models is controlled within 1MB; The system adopts an 8-bit integer quantization method in cooperation with the TensorRT inference acceleration module. The specific formula is as follows: x≈s·(x q -z) where S is the scaling factor, z is the zero point, round(.) is the rounding function, and x q is the quantized integer value, and it is agreed that x q ∈ [0, 255] or x q ∈ [-128, 127]; The knowledge distillation mechanism is used to guide the student model to learn the skeleton action feature distribution by using the teacher model. The distillation loss function is composed of the KL divergence between the output of the student model and the soft label of the teacher model. The specific formula is as follows: Among them, y s is the output of the student model, y is the true label, and z s , z t are the logits of the student and teacher models respectively, T is the temperature coefficient, and α is the weight coefficient.

Images