Cloud edge-end collaborative lightweight acupuncture point recognition method for moxibustion robot

Abstract

The invention provides a cloud edge-end collaborative lightweight acupoint recognition method for a moxibustion robot, and relates to the technical field of acupoint recognition, and the method comprises the steps: collecting a human body image; constructing an acupuncture point detection model which comprises a feature extraction network, a human body context attention module, a dense connection network, an affinity field reasoning module and a confidence degree reasoning module; carrying out key point detection on the acquired human body image based on an acupoint detection model, and extracting human body key point coordinates; and constructing an acupuncture point positioning model, and positioning each acupuncture point based on the human body key point coordinates to obtain coordinate values of the acupuncture points. According to the method, the geometric structure features of the acupuncture points and the traditional Chinese medicine bone degree positioning method are combined, the BCAM is introduced, the prior feature signals are constructed, the acupuncture points of different people in the complex environment can be rapidly and effectively recognized, compared with a previous recognition method, the robustness and generalization are improved, and core support is provided for the accuracy of moxibustion physiotherapy.

Description

Technical Field

[0001] This invention belongs to the field of acupoint recognition technology, and particularly relates to a lightweight acupoint recognition method with cloud-edge-device collaboration for moxibustion robots. Background Technology

[0002] Currently, most moxibustion tasks still rely on manual methods, which are relatively inefficient and hinder the development of traditional Chinese medicine moxibustion. Meanwhile, with the rise of artificial intelligence in recent years, combining traditional Chinese medicine with AI can effectively compensate for the shortcomings of manual moxibustion and greatly improve its efficiency. However, the current mainstream deployment of TCM moxibustion robots in the industry still focuses on "independent device operation," and its deployment methods share the following common shortcomings:

[0003] Most existing moxibustion robots on the market do not have acupoint recognition capabilities. Only a small number of moxibustion robots are equipped with acupoint recognition systems, which are only roughly calculated based on the formula of human body contour. Under conditions such as dynamics, partial occlusion, and changes in lighting, the recognition effect is poor and the robustness is not strong. This solution lacks stability, has low efficiency, and poor calculation accuracy, and is not suitable for large-scale applications.

[0004] Most existing moxibustion robots are deployed entirely locally, relying entirely on their own hardware. This results in high hardware costs, high maintenance costs, and the overall size and weight of the devices make them inconvenient to move. Furthermore, the high integration of individual devices increases potential points of failure, and hardware upgrades require complete replacement, further increasing maintenance costs. These factors significantly limit the large-scale application of moxibustion robots.

[0005] Deep network models require high-performance GPUs / TPUs for accurate recognition, but high-performance hardware is expensive and bulky. Meanwhile, moxibustion robots are limited by the space constraints of physiotherapy rooms and cost control requirements, making it difficult to implement deep network models. Summary of the Invention

[0006] To address the shortcomings of existing technologies, firstly, this invention provides a lightweight acupoint recognition method for moxibustion robots using cloud-edge-device collaboration, comprising the following steps:

[0007] Acquire human images, including the back and abdominal regions;

[0008] An acupoint detection model is constructed, and key points are detected in the collected human images based on the acupoint detection model to extract the coordinates of human key points. The acupoint detection model includes a feature extraction network, a human context attention module, a dense connection network, an affinity field inference module, and a confidence inference module.

[0009] The feature extraction network is used to extract features from human images to obtain preliminary feature map mappings; the human context attention module is used to establish human dependency relationships on the preliminary feature map mappings to obtain enhanced feature maps; the dense connection network is used to generate a set of component affinity fields and a set of confidence maps; the affinity field inference module is used to perform iterative inference on the set of component affinity fields to obtain a predicted set of component affinity fields; the confidence inference module is used to perform iterative inference on the predicted set of component affinity fields and the set of confidence maps to generate a predicted set of confidence maps and extract the coordinates of human key points.

[0010] Acupoint location model is constructed, and each acupoint is located based on the coordinates of key points on the human body to obtain the coordinate values ​​of the acupoints.

[0011] Furthermore, the human context attention module includes oriented convolution and soft spatial attention layers;

[0012] The initial feature map mapping is used to establish human dependency relationships. The specific method is as follows:

[0013] Directional convolution is used to perform a linear transformation on the initial feature map mapping to extract directional features;

[0014] A soft-space attention layer is used to generate a soft-space attention map of directional features. The directional features are then weighted and fused using the soft-space attention map to obtain an enhanced feature map.

[0015] Furthermore, the dense connection network includes multiple 3×3 convolutions;

[0016] The enhanced feature maps are input into a densely connected network to generate a set of component affinity fields and a set of confidence maps. include Each component's affinity field Confidence graph set include Individual confidence plot .

[0017] Furthermore, the affinity field inference module includes multiple convolutional sub-modules connected in series and multiple 1×1 convolutional layers. Each convolutional sub-module includes three 3×3 convolutional layers connected in series. The outputs of the three 3×3 convolutional layers are concatenated to obtain the output of the convolutional sub-module.

[0018] The predicted set of component affinity fields is obtained by iterative reasoning on the component affinity field set. The specific method is as follows:

[0019] Enhanced feature map and component affinity field set Input affinity field inference module for iterative inference, including The first stage , with enhanced feature maps As input to the affinity field inference module, it generates the component affinity field set of the first stage output. ; in the stage , with enhanced feature maps and the The set of component affinity fields output at each stage As input to the affinity field inference module, the first... The set of component affinity fields output at each stage As shown in the formula below:

[0020]

[0021] Finish After several stages of iterative reasoning, the predicted set of component affinity fields is obtained. .

[0022] Furthermore, the series of multiple convolutional sub-modules and multiple 1×1 convolutional layers;

[0023] Iterative reasoning is performed on the predicted set of component affinity fields and the set of confidence maps to generate a set of predicted confidence maps and extract the coordinates of key human body points. The specific method is as follows:

[0024] Set the predicted component affinity fields Input confidence inference module for iterative inference, including The first stage That is, the first module of affinity field inference The stage is based on the predicted component affinity field set. and enhanced feature maps As input to the confidence inference module, it generates a set of confidence maps for the first stage output. ; in the stage , with the predicted component affinity field set Enhanced feature map and the Set of confidence graphs for stage output As input to the confidence inference module, generate Set of confidence graphs for stage output As shown in the formula below:

[0025]

[0026] Finish After several stages of iterative inference, a set of confidence maps for prediction is obtained. .

[0027] Furthermore, the specific method for extracting the coordinates of key human body points is as follows:

[0028] From the set of predicted confidence maps Multiple candidate regions for human body key points are obtained, with each human body key point corresponding to a candidate region. In each candidate region, the key point with the highest confidence value is selected as the human body key point corresponding to that candidate region, and the coordinates of each human body key point are obtained.

[0029] Furthermore, an acupoint localization model was constructed, including a multilayer perceptron.

[0030] The coordinates of each acupoint are located based on the coordinates of key points on the human body, and the coordinate values ​​of the acupoints are obtained. The specific method is as follows:

[0031] The coordinates of key human points are input into the acupoint positioning model as parameters to obtain the relative positions of each key human point and the known reference standard position, generate a set of relative positions, and construct structural set constraints.

[0032] Based on structural set constraints, the positional distance and orientation between each human body key point are calculated, and each human body key point is connected. A structural guidance tensor is generated based on the positional distance and orientation. The structural guidance tensor and the preliminary feature map mapping of the human body image are fused and then subjected to 1×1 convolution to generate prior feature signals.

[0033] Using the longitudinal direction of the torso as a reference, a local coordinate system is established by selecting one key human body point as the origin and another key human body point is selected. The basic offset angle is calculated based on the coordinates of the two key human body points.

[0034] Two human body key points are selected, and the predicted distance between the two human body key points is calculated based on the coordinates of the two human body key points. The theoretical distance between the two human body key points is obtained, and the ratio of the predicted distance to the theoretical distance is used as the bone ratio factor.

[0035] Multilayer perceptron is used to optimize and correct the prior feature signal, basic offset angle, and bone scale factor, generating corrected scale factor and corrected offset angle.

[0036] A reference acupoint is selected, and each acupoint is located based on the corrected scaling factor and the corrected offset angle to obtain the coordinate value of each acupoint. As shown in the formula below:

[0037]

[0038] in, The x-coordinate of the reference acupoint is... The vertical coordinate of the reference acupoint is... This is the corrected scaling factor. This is the corrected offset angle. For the first The x-coordinate of each acupoint x-coordinate relative to the reference acupoint The offset, For the first The vertical coordinate of each acupoint The vertical coordinate relative to the reference acupoint The offset.

[0039] Secondly, the present invention provides a lightweight acupoint recognition system for moxibustion robots with cloud-edge-device collaboration, comprising a terminal layer, an edge computing layer and a cloud layer;

[0040] The terminal layer is used to collect real-time data and user information, as well as to execute moxibustion tasks and provide human-computer interaction functions. Real-time data includes human body images, moxibustion head working status information, and moxibustion robot running posture parameters.

[0041] The edge computing layer is used to preprocess and detect the status of real-time data collected by the terminal layer. It is also used to receive moxibustion plans from the cloud layer, plan moxibustion tasks based on the moxibustion plans, control the terminal layer to execute moxibustion tasks, and monitor the operating status of the terminal layer and the user status.

[0042] The cloud layer is used to locate acupoints and generate moxibustion plans based on preprocessed real-time data. It is also used to store preprocessed real-time data, user information, and an expert acupoint database, which includes acupoint location rules and treatment plans.

[0043] The terminal layer includes several moxibustion robots and a user control terminal. Each moxibustion robot comprises a high-precision servo control system, a depth camera, a moxibustion head, and a distance sensor. The high-precision servo control system controls the robot's movements, the depth camera captures images of the user's body, and the distance sensor collects the robot's pose parameters and the distance between the moxibustion head and the user. The user control terminal includes a data acquisition module and a human-computer interaction module. The data acquisition module obtains user information, body images, the robot's pose parameters, the distance between the moxibustion head and the user, and the moxibustion head's working status information. The user control terminal communicates with both the edge computing layer and the cloud layer, uploading real-time data to the edge computing layer and user information to the cloud layer. The human-computer interaction module provides interactive functions, including allowing the user to select moxibustion techniques, moxibustion time, and acupoints.

[0044] The edge computing layer includes several industrial control computers, each comprising a preprocessing module, a real-time detection module, and a control module. The preprocessing module preprocesses the real-time data. The real-time detection module monitors the status of the preprocessed real-time data, determining whether it meets requirements, including human image clarity detection and whether human posture conforms to preset standards, and uploads the preprocessed real-time data to the cloud layer. The control module plans and controls the moxibustion robot in the terminal layer to perform moxibustion tasks, monitoring the robot's operating status and human posture.

[0045] The cloud layer includes an acupoint detection model, an acupoint location model, and a cloud database. The acupoint detection model is used to extract key points of the human body from human images. The acupoint location model locates the coordinate values ​​of acupoints based on the key points of the human body and generates a moxibustion plan data package based on the acupoint coordinate values. The cloud database is used to store preprocessed real-time data, user information, and an expert acupoint database.

[0046] Thirdly, this application proposes a computer program product, including a computer program or instructions, which, when executed by a processor, implements the cloud-edge-device collaborative lightweight acupoint recognition method for moxibustion robots.

[0047] The beneficial effects of adopting the above technical solution are as follows: The present invention provides a cloud-edge-device collaborative lightweight acupoint recognition method for moxibustion robots, which constructs an acupoint detection model and an acupoint location model, combines the geometric structural features of acupoints with the traditional Chinese medicine bone measurement method, introduces the human body association attention module (BCAM) and constructs prior feature signals, which can quickly and effectively identify acupoints of different people in complex environments. Compared with previous recognition methods, the robustness and generalization are improved, providing core support for the accuracy of moxibustion therapy.

[0048] This invention provides a lightweight acupoint recognition system for moxibustion robots, featuring cloud-edge-device collaboration. Regarding device lightweighting and cost control, the cloud handles high-performance model inference tasks, while the edge computing end only needs to deploy lightweight detection, preprocessing, and control modules. Complex data acquisition and moxibustion task execution on the terminal device do not require integrated high-performance hardware, achieving miniaturization and mobility of the robot, adapting to various physiotherapy scenarios and space requirements. In terms of system stability and scalability, the cloud-edge-device collaborative architecture provides high-performance computing support and model iteration updates through the cloud server, while the edge handles real-time control and local anomaly response. The terminal executes standardized tasks, ensuring system stability and simplifying subsequent maintenance processes. Hardware upgrades only require updating the edge or terminal modules, and software iterations can be completed via cloud push, significantly reducing maintenance costs. Attached Figure Description

[0049] Figure 1A schematic diagram of a lightweight acupoint recognition method for moxibustion robots based on cloud-edge-device collaboration provided in Embodiment 1 of the present invention;

[0050] Figure 2 The overall deployment architecture diagram of the cloud-edge-device collaborative lightweight acupoint recognition system for moxibustion robots provided in Embodiment 2 of the present invention;

[0051] Figure 3 A schematic diagram of the hardware architecture of the moxibustion robot system provided in Embodiment 3 of the present invention;

[0052] Figure 4 The flowchart of the single-unit moxibustion robot system provided in Embodiment 3 of the present invention. Detailed Implementation

[0053] The specific implementation methods of this application will be further described in detail below with reference to the accompanying drawings and embodiments.

[0054] Example 1

[0055] A lightweight acupoint recognition method with cloud-edge-device collaboration for moxibustion robots includes the following steps:

[0056] Acquire human images, including the back and abdominal regions;

[0057] Acupoint detection model is constructed, and key points are detected in the collected human images based on the acupoint detection model to extract the coordinates of key points in the human body.

[0058] Acupoint location model was constructed, and the coordinate values ​​of acupoints were located based on the coordinates of key points on the human body.

[0059] In this embodiment, an acupoint detection model DAug-RCNN (Dual-Augmented RCNN) is constructed, as follows: Figure 1 As shown, it includes a feature extraction network, a Body Context Attention Module (BCAM), a DenseNet dense connection network, an affinity field inference module, and a confidence inference module;

[0060] The acupoint detection model uses a bottom-up keypoint framework as a feature extraction network to preprocess and extract preliminary features from human images; it introduces a human context attention module (BCAM) based on a bidirectional context attention mechanism to extract the geometric dependencies between human keypoints to enhance feature representation; and it generates predicted keypoint coordinates by a series of affinity field inference modules (PAFs) and confidence inference modules (CMDs).

[0061] In this embodiment, the first 10 layers of the VGG-19 network are used as the feature extraction network to extract features from the human image and obtain a preliminary feature map mapping. ;

[0062] Construct a Human Context Attention Module (BCAM), including oriented convolutions and soft-space attention layers; use the BCAM to map the initial feature maps. Human dependency relationships are established to obtain enhanced feature maps. The specific method is as follows:

[0063] Using directional convolution Preliminary feature map mapping Perform linear transformation to extract directional features As shown in the formula below:

[0064]

[0065] Generating directional features using a soft-space attention layer Soft space attention map Using soft space attention maps directional features Weighted fusion yields enhanced feature maps. As shown in the formula below:

[0066]

[0067] in, For the sigmoid function, For convolution operations, This refers to the element-wise multiplication operation between two matrices.

[0068] Enhanced feature map The input consists of a densely connected network, DenseNet, with multiple 3×3 convolutions, generating a set of component affinity fields and a set of confidence maps. include Each component's affinity field Confidence graph set include Individual confidence plot ;

[0069] Construct an affinity field inference module, which includes multiple convolutional sub-modules connected in series and multiple 1×1 convolutional layers. Each convolutional sub-module includes three 3×3 convolutional layers connected in series. The outputs of the three 3×3 convolutional layers are concatenated to obtain the output of the convolutional sub-module.

[0070] Enhanced feature map and component affinity field set Input affinity field inference module for iterative inference, including The first stage , with enhanced feature maps As input to the affinity field inference module, it generates the component affinity field set of the first stage output. ; in the stage , with enhanced feature maps and the Stage-predicted component affinity field set As input to the affinity field inference module, it generates the set of component affinity fields for the output of stage t. As shown in the formula below:

[0071]

[0072] Finish After several stages of iterative reasoning, the predicted set of component affinity fields is obtained. ;

[0073] Construct a confidence inference module, which includes multiple concatenated convolutional sub-modules and multiple 1×1 convolutional layers;

[0074] Set the predicted component affinity fields Input confidence inference module for iterative inference, including The first stage That is, the first module of affinity field inference The stage is based on the predicted component affinity field set. and enhanced feature maps As input to the confidence inference module, it generates a set of confidence maps for the first stage output. ; in the Phase , with the predicted component affinity field set Enhanced feature map and the Set of confidence graphs for stage output As input to the confidence inference module, generate Set of confidence graphs for stage output As shown in the formula below:

[0075]

[0076] go through After several stages, a set of predicted confidence maps is obtained. ;

[0077] From the set of predicted confidence maps The process involves obtaining candidate regions for human body keypoints, with each keypoint corresponding to a specific candidate region. These keypoints include the left and right acromions, neck, mid-hip, C7 vertebra, and navel. Within each candidate region, the keypoint with the highest confidence value is selected as the corresponding human body keypoint. The coordinates of the key points of an individual's body are ;

[0078] In this embodiment, a training supervision mechanism is added to the affinity field inference module and the confidence inference module to prevent gradient explosion, and a loss function for the affinity field inference module is constructed. Loss function of confidence inference module As shown in the formula below:

[0079]

[0080] in, and These are the true vector field labels for the affinity field of the baseline component and the true vector field labels for the baseline confidence map, respectively. This is a penalty function used to avoid correcting the true labeled data. For the predicted coordinates of key human body points, For the first The vector field labels predicted by the affinity field inference module. For the first Vector field labels predicted by the round confidence inference module;

[0081] Loss function based on affinity field inference module Loss function of confidence inference module Construct the total predicted feature point loss function of the acupoint detection model. As shown in the formula below:

[0082]

[0083] Using the total predicted feature point loss function The acupoint detection model is trained to obtain a well-trained acupoint detection model;

[0084] Constructing an acupoint location model, and locating acupoint coordinates based on the coordinates of key points on the human body, the specific method is as follows:

[0085] The coordinates of key human body points are input as parameters into the acupoint location model. Based on the "Ling Shu·Gu Du" standard, the coordinates of each key human body point and known reference standard positions are obtained. The relative positions are used to generate a set of relative positions. Construct structural set constraints As shown in the formula below:

[0086]

[0087] in, For the first Individual key points relative to reference standard position distance, For the first Individual key points relative to reference standard position distance, and Index numbering for key points on the human body;

[0088] Based on structural set constraints Calculate the positional distances and directions between each key point on the human body. Key points of the individual body and the first Positional distance between key points of an individual's body and direction As shown in the formula below:

[0089]

[0090] in, For the first Coordinates of key points on an individual's body For the first Coordinates of key points on an individual's body;

[0091] Connect the key points of each human body, based on their position and distance. and direction Generate structure-guided tensor As shown in the formula below:

[0092]

[0093] in, Here, is a hyperparameter, and exp is an exponential function. This is element-wise multiplication;

[0094] Structure-guided tensor and preliminary feature map mapping After fusion, the signals are then subjected to 1×1 convolution to generate prior feature signals. As shown in the formula below:

[0095]

[0096] Since different postures of the human body will produce a certain offset angle, this embodiment uses the longitudinal direction of the torso as the reference and the coordinates of the key points of the mid-hip as the reference. Establish a local coordinate system with the origin as the origin, and point the positive X-axis towards the key point coordinates of the vertebral column. The Y-axis is perpendicular to the X-axis, and the positive direction of the Y-axis points to the key coordinates of the right acromion. Calculate the base offset angle As shown in the formula below:

[0097]

[0098] Because the "cun" in Traditional Chinese Medicine is a specific unit that differs from international units of measurement, in this embodiment, the theoretical distance is taken as 16 cun, the distance between the left and right acromions. This is based on the coordinates of the left acromion. Coordinates of the right acromion Calculate the bone density ratio factor As shown in the formula below:

[0099]

[0100]

[0101] in, The predicted distances between the left and right acromions;

[0102] Using a multilayer perceptron to process prior feature signals Basic offset angle Bone ratio factor Perform optimization and correction to generate a scaling factor. and the corrected offset angle As shown in the formula below:

[0103]

[0104] in, It is a multilayer perceptron. This is a set of coordinates of key points on the human body. The coordinates of the key points of the left acromion. The coordinates of the key points of the right acromion. The coordinates of the key point of Dazhui (GV14) Bone density scaling factor From a basic perspective;

[0105] Based on the corrected scaling factor and the corrected offset angle Acupoint coordinates obtained from the coordinates of key points on the human body As shown in the formula below:

[0106]

[0107] in, , The x-coordinates are respectively and ordinate Key coordinates relative to the C7 vertebra The offset;

[0108] offset , According to the *Ling Shu* (Spiritual Pivot), the local coordinate offsets of the main acupoints on the back are shown in Table 1. , The coordinates of the acupoints obtained are as follows: Dazhui (p0), Taodao (p1), Shenzhu (p2), Zhongshu (p3), Mingmen (p4), Yaoyangguan (p5), Yaoshu (p6), Dazhu (p7, p8), Feishu (p9, p10), left and right Jueyinshu (p11, p12), left and right Xinshu (p13, p14), left and right Geshu (p15, p16), left and right Ganshu (p17, p18), left and right Danshu (p19, p20), left and right Pishu (p21, p22), left and right Weishu (p23, p24), left and right Shenshu (p25, p26), left and right Dachangshu (p27, p28), left and right Xiaochangshu (p29, p30), left and right Pangguangshu (p31, p32), left and right Ciliao (p33, p34).

[0109] Table 1. Coordinate offset of major acupoints on the back

[0110]

[0111] Example 2

[0112] A lightweight acupoint recognition system with cloud-edge-device collaboration for moxibustion robots is proposed. It adopts a three-terminal collaborative architecture of cloud, edge computing, and terminal. The acupoint recognition network model is deployed in the cloud, and the terminal acquires on-site data. The edge computing terminal preprocesses the on-site data, monitors human posture, and controls the operation of the moxibustion robot. This system overcomes the problems of fixed device location, large footprint, and simple scene for moxibustion robots. It improves the acupoint recognition effect of moxibustion robots, provides accurate and stable acupoint recognition for the moxibustion process, and makes moxibustion robots easier to promote on a large scale, reducing the burden on doctors.

[0113] This embodiment presents a lightweight acupoint recognition system for moxibustion robots, featuring cloud-edge-device collaboration. Figure 2 As shown, it includes the terminal layer, the edge computing layer, and the cloud layer;

[0114] The terminal layer is used to collect real-time data and user information, as well as to execute moxibustion tasks and provide human-computer interaction functions. Real-time data includes human body images, moxibustion head working status information, and moxibustion robot running posture parameters.

[0115] The edge computing layer is used to preprocess and detect the status of real-time data collected by the terminal layer. It is also used to receive moxibustion plans from the cloud layer, plan moxibustion tasks based on the moxibustion plans, control the terminal layer to execute moxibustion tasks, and monitor the operating status of the terminal layer and the user status.

[0116] The cloud layer is used to locate acupoints and generate moxibustion plans based on preprocessed real-time data. It is also used to store preprocessed real-time data, user information, and an expert acupoint database, which includes acupoint location rules and treatment plans.

[0117] The terminal layer includes several moxibustion robots and a user control terminal. Each moxibustion robot comprises a high-precision servo control system, a depth camera, a moxibustion head, and a distance sensor. The high-precision servo control system controls the robot's movements, the depth camera captures images of the user's body, and the distance sensor collects the robot's pose parameters and the distance between the moxibustion head and the user. The user control terminal includes a data acquisition module and a human-computer interaction module. The data acquisition module obtains user information, body images, the robot's pose parameters, the distance between the moxibustion head and the user, and the moxibustion head's working status information. The user control terminal communicates with both the edge computing layer and the cloud layer, uploading real-time data to the edge computing layer and user information to the cloud layer. The human-computer interaction module provides interactive functions, including allowing the user to select moxibustion techniques, moxibustion time, and acupoints.

[0118] The edge computing layer includes several industrial control computers, each comprising a preprocessing module, a real-time detection module, and a control module. The preprocessing module preprocesses the real-time data. The real-time detection module monitors the status of the preprocessed real-time data, determining whether it meets requirements, including human image clarity detection and whether human posture conforms to preset standards, and uploads the preprocessed real-time data to the cloud layer. The control module plans and controls the moxibustion robot in the terminal layer to perform moxibustion tasks, monitoring the robot's operating status and human posture.

[0119] The cloud layer includes an acupoint detection model, an acupoint location model, and a cloud database. The acupoint detection model is used to extract key points of the human body from human images. The acupoint location model locates the coordinate values ​​of acupoints based on the key points of the human body and generates a moxibustion plan data package based on the acupoint coordinate values. The cloud database is used to store preprocessed real-time data, user information, and an expert acupoint database.

[0120] In the terminal layer of this embodiment, the moxibustion robot is controlled by a mobile device equipped with a customized client APP. The mobile device relies on 5G to provide wide-area, high-bandwidth connectivity and uses WiFi 6 as a backup or local area network to achieve high-speed, low-latency local connectivity. It uses dual-mode communication to complete interactive tasks such as uploading user information and downloading solutions with the cloud layer. The moxibustion robot is equipped with a high-precision depth camera, Intel RealSense D435i, which acquires real-time data and transmits the real-time data to the edge computing layer via industrial Ethernet. As a field execution device, the moxibustion robot also uses industrial Ethernet to communicate with the edge computing layer, which can accurately and quickly complete the specified moxibustion tasks.

[0121] In this embodiment, the industrial control computer uses Advantech IPC-610L, a high-performance and high-reliability model, to preprocess and detect human body status of the real-time data collected by the terminal layer. This includes making preliminary judgments on the real-time data, quickly checking whether the human body image meets the required clarity and whether the human body posture is standard, and uploading the data that meets the requirements to the cloud server via 5G / WiFi 6; waiting to receive the moxibustion scheme data package sent by the cloud layer, planning and controlling the moxibustion robot to perform the moxibustion task; acquiring real-time data through the depth camera of the moxibustion robot, monitoring whether the equipment operation changes, generating a human body recognition box to detect the human body status in real time, such as recognizing whether the user is lying flat according to the specifications.

[0122] In this embodiment, the cloud layer utilizes a cloud server cluster. The system is based on a Tencent Cloud HCCPNV4h instance, configured with a high-performance GPU: NVIDIA A100 NVLink 40GB for acupoint recognition model inference, providing high computing power support. The built-in cloud database uses MySQL. The Flask lightweight framework is used to build API interfaces to handle requests from the edge layer and the app, such as user information uploads and image reception. Gunicorn is used to enhance the HTTP server's concurrent processing capabilities and stability. Tencent Cloud employs strict security group rules and firewall policies to effectively protect user information security and system stability. The data flow in the cloud layer is as follows: after receiving data from the edge computing layer and terminal layer information, it is passed to the acupoint recognition model, which generates moxibustion data scheme data based on the expert database and then distributes it to the edge computing layer.

[0123] Example 3

[0124] A deformable scheme for a lightweight, cloud-edge-device collaborative acupoint recognition method for moxibustion robots, wherein the hardware architecture of the moxibustion robot system is as follows: Figure 3As shown, based on the lightweight cloud-edge-device collaborative acupoint recognition method for moxibustion robots provided in Embodiment 1 and the lightweight cloud-edge-device collaborative acupoint recognition system for moxibustion robots provided in Embodiment 2, a single moxibustion robot is controlled to complete the moxibustion task. Figure 4 As shown, it includes the following steps:

[0125] Step 1: The user establishes data communication with the moxibustion robot through the user control terminal equipped with the client, initializes the moxibustion robot through the client, obtains user information, and uploads the user information to the cloud layer; the user information includes age, height and weight, moxibustion duration and moxibustion technique;

[0126] Step 2: After receiving the user information, the cloud layer verifies the IP address of the user's control terminal. Once the verification is successful, the cloud layer stores the user information in the cloud database, generates a unique response code, and sends it to the edge computing layer.

[0127] Step 3: After the edge computing layer receives the response code, the industrial control computer controls the moxibustion robot in the terminal layer to turn on the depth camera and guide the user to lie flat as required through voice prompts. It acquires multiple human images including the back and abdominal areas, including color images and depth images, and uploads them to the industrial control computer for preliminary judgment of whether the human images meet the preset standards, including whether the user is lying flat according to the standard. If the user is not lying flat according to the standard, the corresponding prompt will be issued.

[0128] Step 4: The industrial computer preprocesses the human body images that meet the preset standards, including format adjustment and enhancement of the human body images;

[0129] The human body image that meets the preset standard is adjusted to a fixed resolution of 640×480 through cropping and filling. The aspect ratio of the human body image is maintained and the edges are processed by filling calculation. The human body image after format adjustment is normalized.

[0130] The normalized human body image is subjected to multiple median filters with different filter kernel sizes to remove salt-and-pepper noise and improve the quality of the human body image, resulting in a preprocessed human body image, which is then uploaded to the cloud layer.

[0131] Step 5: Input the preprocessed human body image into the acupoint detection model and acupoint location model for inference calculation to obtain the coordinate data of each acupoint in the detection area. Perform coordinate transformation on each acupoint coordinate data and the corresponding depth data to obtain the three-dimensional spatial coordinates in the base coordinate system of the moxibustion robot.

[0132] Step 5.1: Register the color image and the depth image, and convert the coordinate data of each acupoint obtained from the color image into acupoint coordinates under the depth image through multiple coordinate transformations;

[0133] Due to the different resolutions of color images and depth images, the coordinates of color pixels and depth pixels at the image plane position in the obtained color image and depth image data streams are inconsistent. In this embodiment, the coordinate transformation is implemented using the depth camera's own registration function library.

[0134] Step 5.2: Pre-determine whether the transformed three-dimensional coordinate values ​​meet the requirements, and exclude invalid or abnormal coordinates. Invalid or abnormal coordinates include coordinate values ​​outside the specified physiotherapy area and unreasonable depth information.

[0135] Step 6: Use the acupoint detection model and acupoint location model to perform multiple inferences to obtain the coordinates of the acupoints identified multiple times. By comparison and filtering, remove the acupoint coordinates with identification accuracy lower than the preset threshold to improve the accuracy of acupoint coordinate identification and obtain the first acupoint coordinates.

[0136] In this embodiment, the mean square error screening method is used to remove acupoint coordinates with large errors, and the average coordinates of acupoint coordinates identified multiple times are calculated. As shown in the formula below:

[0137]

[0138] in, Let x be the x-coordinate of the k-th acupoint obtained from the i-th identification. Let y be the ordinate of the k-th acupoint obtained from the i-th identification, and N be the number of inferences;

[0139] The root mean square error of the coordinates of each acupoint is calculated using the following formula:

[0140]

[0141] Set an identification error threshold T, discard acupoints with errors exceeding the threshold T, repeatedly calculate the average coordinates of acupoints, and select the coordinates of the acupoint with the smallest mean square error as the coordinates of the first acupoint.

[0142] Step 7: Based on the coordinates of the first acupoint, generate a moxibustion plan according to the plans in the expert database, including the acupoints, moxibustion time at each point, distance between the moxibustion head and the skin, and technique information;

[0143] Step 8: The industrial control computer downloads the moxibustion plan from the cloud and controls the moxibustion robot to perform the moxibustion task;

[0144] The moxibustion robot's moxibustion head is controlled using a vertical positioning method to hover above the target acupoint. The plane at the tip of the moxibustion head is parallel to the skin plane. The normal vector coordinates of the moxibustion head tip are calculated based on depth images. According to the normal vector coordinates of the moxibustion head tip The system calculates the spatial coordinates from the current position of the moxibustion head to the target acupoint and the trajectory of the center point of rotation. It adjusts the suspension distance and movement speed of the moxibustion head in real time through a distance sensor and executes the moxibustion task according to the moxibustion technique selected by the user.

[0145] The moxibustion robot is controlled to locate acupoints in the human body image area one by one according to the moxibustion plan. After one acupoint is completed, the moxibustion robot is controlled to perform moxibustion on the next target acupoint according to the moxibustion plan until the entire moxibustion process is completed. When the moxibustion robot is performing the moxibustion task, the operating status of the moxibustion robot and the user's human posture are monitored in real time by the industrial control computer to ensure the safe operation of the moxibustion robot.

[0146] Example 4

[0147] A modifiable scheme for a lightweight acupoint recognition method with cloud-edge-device collaboration for moxibustion robots, based on the lightweight acupoint recognition method with cloud-edge-device collaboration for moxibustion robots provided in Embodiment 1 and the lightweight acupoint recognition system with cloud-edge-device collaboration for moxibustion robots provided in Embodiment 2, controls multiple moxibustion robots to complete moxibustion tasks, including the following steps:

[0148] S1: Install n moxibustion robots at the target site. Each moxibustion robot is equipped with a corresponding industrial control computer. At the same time, the client completes the identity verification of the moxibustion robot on the cloud server and initializes the moxibustion robot.

[0149] S2: After the industrial control computer starts up, it performs a self-test, including: checking the computing power output of the built-in graphics card, the speed of the industrial Ethernet port, and whether the depth camera can start normally; after the self-test passes, it sends a waiting task signal to the cloud; if a fault is detected, it issues an alarm signal.

[0150] S3: When multiple moxibustion robots provide parallel moxibustion services to users, the user control terminal corresponding to each moxibustion robot collects user information, including the user's height, weight, age, physical condition, and past moxibustion data. The user information collected by multiple moxibustion robots is integrated and uploaded from the terminal layer to the cloud database of the cloud layer through 5G / WiFi6 dual-mode communication. The cloud layer generates a unique user ID and associates it with the industrial control computer ID.

[0151] S4: The cloud layer queues the requests from the user control terminals corresponding to each moxibustion robot and generates a unique ID to respond to them. It calls the expert acupoint database to generate a moxibustion plan based on the collected user information. The moxibustion plan includes the user's moxibustion method, the corresponding acupoints, and the moxibustion time for each acupoint.

[0152] S5: After initialization, the moxibustion robot receives a response from the cloud server and guides the user to lie down in the corresponding moxibustion area for preliminary preparation. The industrial control computer controls the Intel RealSense D435i depth camera on the moxibustion robot to collect human images in real time, detect whether the user's posture is correct, such as whether the user's torso is overly tilted, and repeatedly guides the user through voice until the user's posture meets the preset standard.

[0153] S6: Acquire human body images and perform acupoint recognition, detecting all acupoints of the user, including the following steps:

[0154] S6.1: The industrial control computer controls the moxibustion robot to open the depth camera and collect images of the user's target area, acquiring the user's human body image, including color and depth images; the industrial control computer detects whether the user's posture meets the preset standards based on the human body image, and transmits the human body image that meets the preset standards to the preprocessing module. The preprocessing module crops the human body image to 640×480, normalizes the pixels to the range of [0,1], removes noise in the human body image through median filtering, and retains the human body image feature information; the preprocessed human body image is then uploaded to the cloud layer.

[0155] S6.2: The cloud layer obtains the industrial control computer ID and the corresponding pre-processed human body image, inputs the pre-processed human body image into the acupoint detection model and the acupoint positioning model for inference calculation, calls the acupoint detection model to generate the coordinates of human body key points, calls the acupoint positioning model, and locates the acupoint coordinate values ​​based on the human body key point coordinates.

[0156] In this embodiment, taking the Taodao acupoint as an example: the coordinates of key points on the user's body, including the left acromion, are obtained through the acupoint detection model. right acromion Mid-hip Dazhui Then, the optimized proportions were obtained using an acupoint location model. Offset angle And calculate the coordinates of each acupoint. As shown in the formula below:

[0157]

[0158] Based on the acupoint reference table, the key coordinates of the Taodao acupoint relative to the Dazhui acupoint were obtained. offset Further refine the coordinates of the Taodao acupoint. for:

[0159]

[0160] S6.3: After obtaining the acupoint coordinates corresponding to several images through cloud-based inference, the acupoint coordinates are filtered to obtain high-precision acupoint coordinates that conform to the corresponding moxibustion method.

[0161] For the The coordinates of acupoints in the frame of human body image are filtered and calculated. The average coordinates of acupoints in a frame of human image are given by the following formula:

[0162]

[0163] calculate The mean square error of acupoint coordinates in a frame of human image is shown in the following formula:

[0164]

[0165] Setting the accuracy threshold for moxibustion By eliminating abnormal points, high-precision acupoint coordinates that meet the requirements of moxibustion are obtained; in this embodiment, a moxibustion precision threshold is set. ;

[0166] S6.4: The industrial control computer obtains high-precision acupoint coordinates and performs the same calculation on the depth image to avoid error estimation caused by the instability of the depth camera detection, and obtains the depth coordinates corresponding to all the user's acupoint coordinates.

[0167] S6.5: Set the coordinates of each acupoint. By associating these coordinates with the corresponding depth coordinates, the actual coordinates of each acupoint in the camera coordinate system can be obtained. The actual coordinates of each acupoint are transformed, converting the initial pose of the moxibustion robot into a homogeneous matrix. The spatial coordinates of each acupoint are then calculated based on the spatial relationship between the camera coordinate system and the moxibustion robot's base coordinate system; that is, the acupoint coordinates in the moxibustion robot's base coordinate system. ;

[0168] S7: Based on the coordinates of each acupoint in the base coordinate system of the moxibustion robot, user information, and the scheme in the expert database, the trajectory of the moxibustion head of the moxibustion robot is planned, and the trajectory of each moxibustion head is sent to the corresponding industrial control computer.

[0169] A list of moxibustion sequences is generated based on plans from the expert database to plan the paths between adjacent meridians and different locations within the same meridian.

[0170] For example, a moxibustion plan might require moxibustion on acupoint A first, followed by acupoint B. Even though the current acupoint is close to acupoint B, the treatment order still needs to be arranged according to Traditional Chinese Medicine (TCM) theory. Therefore, different orders affect the path and trajectory planning of the moxibustion. The order of acupoints is determined according to the moxibustion sequence requirements in TCM theory; if there are no specific order requirements, the shortest path strategy is followed for moxibustion head trajectory planning.

[0171] S9: The industrial control computer executes the moxibustion task according to the planned trajectory, and monitors the operation status of the moxibustion robot and the user's physical condition through the depth camera of the moxibustion robot.

[0172] S9.1: The moxibustion head of the moxibustion robot reaches the target acupoint and is vertically aligned with it. According to the trajectory in the moxibustion plan and the set distance between the moxibustion head and the skin surface, it ignites and performs moxibustion, such as sparrow-pecking moxibustion, rotary moxibustion, etc.

[0173] S9.2: The moxibustion robot repeats S9.1 until the moxibustion task of the current target acupoint is completed. After the moxibustion task of each target acupoint is completed, the user's human body image after moxibustion is uploaded to the cloud layer, compared with the initial human body image, and the user's moxibustion record is updated. After the moxibustion robot completes the moxibustion task of all target acupoints, the moxibustion robot returns to the initial position, the cloud layer generates a moxibustion report and sends it to the user control terminal.

[0174] S9.3: If the moxibustion robot detects an emergency stop signal during the moxibustion task, the moxibustion robot will immediately stop working, return to its initial position, and prompt the user with the reason for the abnormality to ensure the safety and reliability of the equipment. Emergency stop signals include the user pressing the emergency stop button, the user being outside the designated moxibustion area, or insufficient consumables, such as the moxibustion stick length being <3cm.

[0175] Example 5:

[0176] This embodiment proposes an electronic device, including: one or more processors, and a memory, wherein the memory is used to store instructions, and when the instructions are executed by the one or more processors, the one or more processors execute the cloud-edge-device collaborative lightweight acupoint recognition method for moxibustion robots.

[0177] The electronic device can be a mobile phone, computer, or tablet computer, etc., and includes a memory and a processor. The memory stores a computer program, which, when executed by the processor, implements the cloud-edge-device collaborative lightweight acupoint recognition method for moxibustion robots as described in the embodiments. It is understood that the electronic device may also include input / output (I / O) interfaces and communication components.

[0178] The processor is used to execute all or part of the steps in the cloud-edge-device collaborative lightweight acupoint recognition method for moxibustion robots as described in the above embodiments. The memory is used to store various types of data, which may include, for example, instructions for any application or method in the electronic device, as well as application-related data.

[0179] The processor can be implemented as an Application Specific Integrated Circuit (ASIC), Digital Signal Processor (DSP), Programmable Logic Device (PLD), Field Programmable Gate Array (FPGA), controller, microcontroller, microprocessor, or other electronic components, and is used to execute the cloud-edge-device collaborative lightweight acupoint recognition method for moxibustion robots described in the above embodiments.

[0180] Example 6:

[0181] This embodiment proposes a computer-readable storage medium that stores executable instructions. When these instructions are executed, if they are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium.

[0182] The computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the cloud-edge-device collaborative lightweight acupoint recognition method for moxibustion robots described in the various embodiments of this application.

[0183] The aforementioned storage media include: flash memory, hard disk, multimedia card, card-type memory (e.g., SD (Secure Digital Memory Card) or DX (Memory Data Register, MDR) memory, etc.), random access memory (RAM), static random-access memory (SRAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), programmable read-only memory (PROM), magnetic storage, disk, optical disk, server, APP (Application) application store, and other media capable of storing program verification codes. These media store computer programs, which, when executed by a processor, can implement the various steps of the cloud-edge-device collaborative lightweight acupoint recognition method for moxibustion robots described above.

[0184] Example 7:

[0185] This embodiment proposes a computer program product, including a computer program or instructions, which, when executed by a processor, implements the cloud-edge-device collaborative lightweight acupoint recognition method for moxibustion robots.

[0186] Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or part of the technical solution, can be embodied in the form of a computer program product.

[0187] The various embodiments in this application are described in a progressive manner. The same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on describing the differences from other embodiments.

[0188] The scope of protection of this application is not limited to the embodiments described above. Obviously, those skilled in the art can make various modifications and variations to this disclosure without departing from the scope and spirit of this disclosure. If such modifications and variations fall within the scope of this disclosure and its equivalents, then the intent of this disclosure also includes these modifications and variations.

Claims

1. A lightweight acupoint recognition method for cloud-edge-device collaboration for moxibustion robots, characterized in that, Includes the following steps: Acquire human images, including the back and abdominal regions; An acupoint detection model is constructed, and key points are detected in the collected human images based on the acupoint detection model to extract the coordinates of human key points. The acupoint detection model includes a feature extraction network, a human context attention module, a dense connection network, an affinity field inference module, and a confidence inference module. The feature extraction network is used to extract features from human images to obtain preliminary feature map mappings; the human context attention module is used to establish human dependency relationships on the preliminary feature map mappings to obtain enhanced feature maps; the dense connection network is used to generate a set of component affinity fields and a set of confidence maps; the affinity field inference module is used to perform iterative inference on the set of component affinity fields to obtain a predicted set of component affinity fields; the confidence inference module is used to perform iterative inference on the predicted set of component affinity fields and the set of confidence maps to generate a predicted set of confidence maps and extract the coordinates of human key points. Acupoint location model is constructed, and each acupoint is located based on the coordinates of key points on the human body to obtain the coordinate values ​​of the acupoints.

2. The lightweight acupoint recognition method for cloud-edge-device collaboration for moxibustion robots according to claim 1, characterized in that, The human context attention module includes oriented convolution and soft spatial attention layers; The initial feature map mapping is used to establish human dependency relationships. The specific method is as follows: Directional convolution is used to perform a linear transformation on the initial feature map mapping to extract directional features; A soft-space attention layer is used to generate a soft-space attention map of directional features. The directional features are then weighted and fused using the soft-space attention map to obtain an enhanced feature map.

3. A lightweight acupoint recognition method for cloud-edge-device collaboration for moxibustion robots according to claim 2, characterized in that, The dense connection network includes multiple 3×3 convolutions; The enhanced feature maps are input into a densely connected network to generate a set of component affinity fields and a set of confidence maps. include Each component's affinity field Confidence graph set include Individual confidence plot .

4. A lightweight acupoint recognition method for cloud-edge-device collaboration for moxibustion robots according to claim 3, characterized in that, The affinity field inference module includes multiple convolutional sub-modules connected in series and multiple 1×1 convolutional layers. Each convolutional sub-module includes three 3×3 convolutional layers connected in series. The outputs of the three 3×3 convolutional layers are concatenated to obtain the output of the convolutional sub-module. The predicted set of component affinity fields is obtained by iterative reasoning on the component affinity field set. The specific method is as follows: Enhanced feature map and component affinity field set Input affinity field inference module for iterative inference, including The first stage , with enhanced feature maps As input to the affinity field inference module, it generates the component affinity field set of the first stage output. ; in the stage , with enhanced feature maps and the The set of component affinity fields output at each stage As input to the affinity field inference module, the first... The set of component affinity fields output at each stage As shown in the formula below: Finish After several stages of iterative reasoning, the predicted set of component affinity fields is obtained. .

5. A lightweight acupoint recognition method for cloud-edge-device collaboration for moxibustion robots according to claim 4, characterized in that, The series of multiple convolutional sub-modules and multiple 1×1 convolutional layers; Iterative reasoning is performed on the predicted set of component affinity fields and the set of confidence maps to generate a set of predicted confidence maps and extract the coordinates of key human body points. The specific method is as follows: Set the predicted component affinity fields Input confidence inference module for iterative inference, including The first stage That is, the first module of affinity field inference The stage is based on the predicted component affinity field set. and enhanced feature maps As input to the confidence inference module, it generates a set of confidence maps for the first stage output. ; in the stage , with the predicted component affinity field set Enhanced feature map and the Set of confidence graphs for stage output As input to the confidence inference module, generate Set of confidence graphs for stage output As shown in the formula below: Finish After several stages of iterative inference, a set of confidence maps for prediction is obtained. .

6. A lightweight acupoint recognition method for cloud-edge-device collaboration for moxibustion robots according to claim 5, characterized in that, The specific method for extracting the coordinates of key points on the human body is as follows: From the set of predicted confidence maps Multiple candidate regions for human body key points are obtained, with each human body key point corresponding to a candidate region. In each candidate region, the key point with the highest confidence value is selected as the human body key point corresponding to that candidate region, and the coordinates of each human body key point are obtained.

7. A lightweight acupoint recognition method for cloud-edge-device collaboration for moxibustion robots according to claim 6, characterized in that, Construct an acupoint localization model, including a multilayer sensor; The coordinates of each acupoint are located based on the coordinates of key points on the human body, and the coordinate values ​​of the acupoints are obtained. The specific method is as follows: The coordinates of key human points are input into the acupoint positioning model as parameters to obtain the relative positions of each key human point and the known reference standard position, generate a set of relative positions, and construct structural set constraints. Based on structural set constraints, the positional distance and direction between each key point of the human body are calculated, the key points of the human body are connected, and a structural guidance tensor is generated according to the positional distance and direction. The structure-guided tensor and the preliminary feature map mapping of the human image are fused together and then convolved with 1×1 to generate the prior feature signal. Using the longitudinal direction of the torso as a reference, a local coordinate system is established by selecting one key human body point as the origin and another key human body point is selected. The basic offset angle is calculated based on the coordinates of the two key human body points. Two human body key points are selected, and the predicted distance between the two human body key points is calculated based on the coordinates of the two human body key points. The theoretical distance between the two human body key points is obtained, and the ratio of the predicted distance to the theoretical distance is used as the bone ratio factor. Multilayer perceptron is used to optimize and correct the prior feature signal, basic offset angle, and bone scale factor, generating corrected scale factor and corrected offset angle. A reference acupoint is selected, and each acupoint is located based on the corrected scaling factor and the corrected offset angle to obtain the coordinate value of each acupoint. As shown in the formula below: in, The x-coordinate of the reference acupoint is... The vertical coordinate of the reference acupoint is... This is the corrected scaling factor. This is the corrected offset angle. For the first The x-coordinate of each acupoint x-coordinate relative to the reference acupoint The offset, For the first The vertical coordinate of each acupoint The vertical coordinate relative to the reference acupoint The offset.

8. A lightweight acupoint recognition system with cloud-edge-device collaboration for moxibustion robots, characterized in that, It includes the terminal layer, the edge computing layer, and the cloud layer; The terminal layer is used to collect real-time data and user information, as well as to execute moxibustion tasks and provide human-computer interaction functions. Real-time data includes human body images, moxibustion head working status information, and moxibustion robot running posture parameters. The edge computing layer is used to preprocess and detect the status of real-time data collected by the terminal layer. It is also used to receive moxibustion plans from the cloud layer, plan moxibustion tasks based on the moxibustion plans, control the terminal layer to execute moxibustion tasks, and monitor the operating status of the terminal layer and the user status. The cloud layer is used to locate acupoints and generate moxibustion plans based on preprocessed real-time data. It is also used to store preprocessed real-time data, user information, and an expert acupoint database, which includes acupoint location rules and treatment plans.

9. A lightweight acupoint recognition system for moxibustion robots with cloud-edge-device collaboration according to claim 8, characterized in that, The terminal layer includes several moxibustion robots and a user control terminal. Each moxibustion robot comprises a high-precision servo control system, a depth camera, a moxibustion head, and a distance sensor. The high-precision servo control system controls the robot's movements, the depth camera captures images of the user's body, and the distance sensor collects the robot's pose parameters and the distance between the moxibustion head and the user. The user control terminal includes a data acquisition module and a human-computer interaction module. The data acquisition module obtains user information, body images, the robot's pose parameters, the distance between the moxibustion head and the user, and the moxibustion head's working status information. The user control terminal communicates with both the edge computing layer and the cloud layer, uploading real-time data to the edge computing layer and user information to the cloud layer. The human-computer interaction module provides human-computer interaction functions, including selecting moxibustion techniques, moxibustion time, and acupoints. The edge computing layer includes several industrial control computers, which include a preprocessing module, a real-time detection module, and a control module. The preprocessing module is used to preprocess real-time data; The real-time monitoring module is used to monitor the status of the pre-processed real-time data and determine whether the pre-processed real-time data meets the requirements, including human image clarity detection and whether human posture meets the preset standards, and uploads the pre-processed real-time data to the cloud layer; the control module is used to plan and control the moxibustion robot in the terminal layer to perform moxibustion tasks and monitor the operating status of the moxibustion robot and human posture. The cloud layer includes an acupoint detection model, an acupoint location model, and a cloud database. The acupoint detection model is used to extract key points of the human body from human images. The acupoint location model locates the coordinate values ​​of acupoints based on the key points of the human body and generates a moxibustion plan data package based on the acupoint coordinate values. The cloud database is used to store pre-processed real-time data, user information, and expert acupoint databases.

10. A computer program product, characterized in that, When the computer program is executed by the processor, it implements the cloud-edge-device collaborative lightweight acupoint recognition method for moxibustion robots as described in any one of claims 1-7.

Images