Children massage acupoint recognition method and system based on artificial intelligence

Abstract

The invention discloses a child massage acupoint recognition method and system based on artificial intelligence, and relates to the technical field of computer vision, and the method comprises the steps: collecting child massage data, and carrying out the preprocessing; inputting the video stream image data into a pre-trained child posture estimation model, and outputting joint point coordinates; identifying a limb movement track of the child, and outputting an optimized joint point coordinate set; based on the optimized joint point coordinate set, according to a traditional Chinese medicine acupoint mapping rule, performing rule mapping on the child target acupoints, and outputting massage acupoint coordinates; constructing an acupoint topological relation graph according to the massage acupoint coordinates and the optimized joint point coordinate set, and outputting a high-precision acupoint coordinate set; and superposing the high-precision acupoint coordinate set after the three-dimensional coordinate conversion and the acupoint action data into video stream image data to generate an acupoint visual interface. According to the invention, real-time accurate positioning and self-adaptive calibration of massage acupoints of children are realized through double-mechanism cooperation of dynamic motion compensation and traditional Chinese medicine topological optimization.

Description

Technical Field

[0001] This invention relates to the field of computer vision technology, and in particular to a method and system for recognizing acupoints for children's massage based on artificial intelligence. Background Technology

[0002] The technology for recognizing acupoints in pediatric massage is rapidly evolving towards a fusion of multimodal perception and intelligent decision-making. Current mainstream solutions rely on high-resolution visual sensors to collect children's posture data, combining this with traditional Chinese medicine acupoint mapping rules to construct a multi-source data system. Lightweight convolutional neural networks (CNNs) are used to achieve posture estimation and preliminary acupoint localization. Sparse optical flow algorithms are gradually being applied to dynamic posture tracking, spatiotemporal displacement matrices provide a basic framework for motion compensation, and edge computing terminal deployments significantly improve real-time performance. Multimodal feature fusion algorithms are driving acupoint recognition from static coordinate extraction to dynamic trajectory modeling, and graph neural networks (GNNs) are gradually incorporating spatial relationship reasoning, providing a technological foundation for precise operation in traditional Chinese massage.

[0003] However, existing technologies have limitations in terms of dynamic tracking accuracy and the depth of integration with traditional Chinese medicine rules. Traditional solutions mostly rely on static joint detection of single-frame images, without establishing motion vector field modeling between consecutive frames. This makes it difficult to compensate for acupoint offsets caused by limb displacement in children, and the coordinate confidence of the output of lightweight models is insufficient. The combination of traditional Chinese medicine acupoint mapping rules and AI models is still at the level of surface matching, lacking a high-order spatial reasoning mechanism based on topological relationships, which makes acupoint localization susceptible to individual body size differences. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides an artificial intelligence-based method for recognizing acupoints for children's massage, which solves the limitations in dynamic tracking accuracy and the depth of integration with traditional Chinese medicine rules.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: In a first aspect, the present invention provides an artificial intelligence-based method for identifying acupoints in children's massage, comprising: collecting children's massage data and preprocessing it; the children's massage data includes video stream image data, traditional Chinese medicine acupoint mapping rules, and acupoint effect data; inputting the video stream image data into a pre-trained children's posture estimation model, the children's posture estimation model performing coordinate decoding on the video stream image data and outputting joint coordinates; extracting the motion vector field between consecutive frames of the video stream image data using a sparse optical flow algorithm, identifying the trajectory of children's limb movements, generating a spatiotemporal displacement matrix, compensating and fusing the joint coordinates with the spatiotemporal displacement matrix, and outputting an optimized joint coordinate set; based on the optimized joint coordinate set, performing regular mapping on the children's target acupoints according to the traditional Chinese medicine acupoint mapping rules, and outputting massage acupoint coordinates; constructing an acupoint topology map based on the massage acupoint coordinates and the optimized joint coordinate set, using a graph convolutional neural network to perform spatial relationship reasoning on the acupoint topology map, and outputting a high-precision acupoint coordinate set; performing three-dimensional coordinate transformation on the high-precision acupoint coordinate set, and superimposing the three-dimensional coordinate transformed high-precision acupoint coordinate set and acupoint effect data onto the video stream image data to generate an acupoint visualization interface.

[0007] As a preferred embodiment of the artificial intelligence-based acupoint recognition method for children's massage described in this invention, the steps include: inputting video stream image data into a pre-trained child pose estimation model; the child pose estimation model performing coordinate decoding on the video stream image data and outputting joint coordinates; and the specific steps are as follows. The video stream image data is input into a pre-trained child pose estimation model. The child pose estimation model extracts features from the video stream image data through a convolutional neural network layer to generate multi-scale feature maps. Deconvolution and probabilistic activation are performed on the multi-scale feature map to output the probability distribution of key points; Perform peak coordinate analysis on the probability distribution of key points and output the key point coordinates.

[0008] As a preferred embodiment of the artificial intelligence-based children's massage acupoint recognition method of the present invention, the pre-trained children's pose estimation model refers to the collection of children's human body dataset, the annotation of children-specific joint points in the children's human body dataset, and the input of the annotated children's human body dataset into the HRNet-W16 network architecture for transfer learning.

[0009] As a preferred embodiment of the artificial intelligence-based acupoint recognition method for children's massage described in this invention, the steps of extracting the motion vector field between consecutive frames of video stream image data using a sparse optical flow algorithm, recognizing the trajectory of children's limb movements, and generating a spatiotemporal displacement matrix are as follows: Feature point detection is performed on gradient regions in video stream image data, gradient coordinate sets are output, and pixel displacement vector statistics are performed on the gradient coordinate sets of adjacent frames to generate motion vector fields. The motion direction and speed of the child's limbs in the motion vector field are correlated and smoothed using a temporal clustering algorithm, and the motion trajectory of the child's limbs is output. The trajectory of children's limb movements is encoded according to the time sequence of video stream image data to generate a spatiotemporal displacement matrix.

[0010] As a preferred embodiment of the artificial intelligence-based acupoint recognition method for children's massage described in this invention, the specific steps of compensating and fusing the joint coordinates with the spatiotemporal displacement matrix to output an optimized joint coordinate set are as follows: The coordinates of the joints are matched point by point with the spatiotemporal displacement matrix to generate joint-displacement data pairs. Based on the motion vector data in the joint-displacement data pair, the joint coordinates are corrected for displacement, and the compensated coordinates are output. The compensated coordinates and the joint coordinates are weighted, averaged, and encapsulated to output an optimized set of joint coordinates.

[0011] As a preferred embodiment of the artificial intelligence-based children's massage acupoint recognition method of the present invention, the specific steps are as follows: Based on an optimized set of joint point coordinates, the target acupoints of the child are mapped according to traditional Chinese medicine acupoint mapping rules, and the coordinates of the massage acupoints are output. Using the mapping rules of traditional Chinese medicine acupoints, the target acupoints are associated and filtered by optimizing the coordinate set of joint points, and the coordinates of anatomical landmarks are output. Based on the relative position description of different acupoints in the TCM acupoint mapping rules, the offset direction and acupoint distance of the target acupoint relative to the coordinates of the anatomical landmark are determined, and the offset direction and acupoint distance are integrated to generate acupoint offset data; The coordinates of anatomical landmarks and acupoint offset data are combined and validated against the child's body surface boundary to output the coordinates of massage acupoints.

[0012] As a preferred embodiment of the artificial intelligence-based acupoint recognition method for children's massage described in this invention, the steps include: constructing an acupoint topological relationship map based on acupoint coordinates and an optimized set of joint point coordinates; using a graph convolutional neural network to perform spatial relationship reasoning on the acupoint topological relationship map; and outputting a high-precision acupoint coordinate set. The coordinates of massage acupoints and the optimized set of joint point coordinates are used as acupoint nodes, and node association edges between acupoint nodes are constructed according to the mapping rules of traditional Chinese medicine acupoints to generate an acupoint topology graph. The acupoint topology map is input into a graph convolutional neural network, and the neighborhood features of each acupoint node are aggregated through a multi-layer graph attention mechanism to generate acupoint feature vectors. Spatial calibration of acupoint coordinates is performed based on acupoint feature vectors to output a high-precision acupoint coordinate set.

[0013] As a preferred embodiment of the artificial intelligence-based acupoint recognition method for children's massage described in this invention, the steps of performing three-dimensional coordinate transformation on the high-precision acupoint coordinate set, and superimposing the transformed high-precision acupoint coordinate set and acupoint effect data onto the video stream image data to generate an acupoint visualization interface are as follows: A three-dimensional coordinate transformation is performed on the high-precision acupoint coordinate set to generate a spatial acupoint coordinate set, and the spatial acupoint coordinate set is associated with the acupoint effect data to generate acupoint association data; The acupoint-related data is overlaid onto the video stream image data, and the overlaid video stream image data is visually marked, rendered, and interactively added to generate an acupoint visualization interface.

[0014] As a preferred embodiment of the artificial intelligence-based children's massage acupoint recognition method of the present invention, the visual mark rendering and interactive addition refer to drawing pulsating halo marks at the coordinate positions of acupoint association data, adding a transparent touch response layer to the pulsating halo marks, and associating the transparent touch response layer with pre-recorded children's acupoint explanation audio to generate an acupoint visualization interface.

[0015] Secondly, the present invention provides an artificial intelligence-based acupoint recognition system for children's massage, comprising, The data acquisition module collects children's massage data and performs preprocessing; the children's massage data includes video stream image data, traditional Chinese medicine acupoint mapping rules, and acupoint effect data; The joint module is used to input video stream image data into a pre-trained child pose estimation model. The child pose estimation model performs coordinate decoding on the video stream image data and outputs the coordinates of the joint points. The optimization module is used to extract the motion vector field between consecutive frames of video stream image data through the sparse optical flow algorithm, identify the trajectory of children's limb movements, generate a spatiotemporal displacement matrix, compensate and fuse the joint coordinates with the spatiotemporal displacement matrix, and output an optimized set of joint coordinates. The acupoint module is used to perform rule mapping on target acupoints for children based on an optimized set of joint coordinates and according to the rules of traditional Chinese medicine acupoint mapping, and output the coordinates of the massage acupoints. The graph module is used to construct an acupoint topology graph based on the coordinates of massage acupoints and an optimized set of joint point coordinates. It uses a graph convolutional neural network to perform spatial relationship reasoning on the acupoint topology graph and outputs a high-precision set of acupoint coordinates. The interface module is used to perform three-dimensional coordinate transformation on the high-precision acupoint coordinate set, and then overlay the high-precision acupoint coordinate set and acupoint effect data into the video stream image data to generate an acupoint visualization interface.

[0016] The beneficial effects of this invention are as follows: By synergistically employing dynamic motion compensation and traditional Chinese medicine topology optimization mechanisms, real-time accurate positioning and adaptive calibration of acupoints for children's massage are achieved. A sparse optical flow algorithm is used to extract the motion vector field and compensate for and fuse it with the joint coordinates, outputting an optimized set of joint coordinates. This dynamically tracks the trajectory of children's limb movements and eliminates posture offset errors in real time, ensuring continuous accuracy of acupoint positioning in dynamic scenarios. Furthermore, a graph convolutional neural network is used to perform spatial relationship reasoning on the acupoint topology map, modeling anatomical constraints and spatial association rules between acupoints. Based on neighborhood feature aggregation, the spatial coordinate distribution of acupoints is adaptively calibrated, reducing acupoint mapping deviations caused by differences in children's growth and improving the robustness of positioning under complex body shapes. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart of an AI-based method for recognizing acupoints during children's massage.

[0019] Figure 2 This is a schematic diagram of an AI-based acupoint recognition system for children's massage.

[0020] Figure 3 A flowchart for optimizing the generation of joint coordinate sets.

[0021] Figure 4 This is a flowchart of acupoint mapping and topology optimization. Detailed Implementation

[0022] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0023] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0024] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0025] Reference Figures 1-4 This is one embodiment of the present invention, which provides an artificial intelligence-based method for recognizing acupoints for children's massage, comprising the following steps: S1. Collect children's massage data and perform preprocessing.

[0026] Children's massage data includes video stream image data, traditional Chinese medicine acupoint mapping rules, and acupoint function data; Specifically, a fixed RGB camera device is used to capture real-time video streams of the child's massage area. The RGB camera device collects video stream image data at a fixed frame rate (e.g., 30 frames per second). The video stream image data includes dynamic images of the human massage area, and the dynamic images cover the anatomical area where the target acupoint is located. Traditional Chinese medicine acupoint mapping rules and acupoint function data are extracted from publicly available literature on children's acupoint location standard maps. The traditional Chinese medicine acupoint mapping rules include acupoint names, names of associated anatomical landmarks, descriptions of relative position and direction, and proportional parameters of body-scale distance. The acupoint function data includes acupoint function classification (e.g., regulating the spleen and stomach and unblocking meridians), associated symptoms, and operation parameters (e.g., massage intensity and frequency).

[0027] Preprocessing includes denoising, resolution normalization, and frame rate synchronization of video stream image data; structured encoding of TCM acupoint mapping rules; and structured encapsulation and semantic encoding adaptation of acupoint effect data. Specifically, multi-frame temporal denoising is performed on the video stream image data acquired by the fixed RGB camera device. A non-local mean filtering algorithm is used to eliminate dynamic motion artifacts and random noise. The non-local mean filtering algorithm performs a weighted average of pixel values ​​based on image block similarity statistics, outputting denoised video stream image data. The denoised video stream image data is then subjected to resolution standardization. A bilinear interpolation algorithm is used to uniformly scale the image size to a fixed pixel resolution (e.g., 512×512 pixel resolution) to ensure that all input frames have the same spatial size. Finally, the frame rate of the resolution-standardized video stream image data is synchronized using a timestamp alignment mechanism to match and stabilize the video stream frame rate at 30 frames per second, forming standardized video stream image data. The mapping relationship between acupoint names and associated anatomical landmark names is extracted from the TCM acupoint mapping rules, generating a binary node connection matrix. The row index of the matrix represents the acupoint name, and the column index represents the anatomical landmark name. Simultaneously, the relative position direction description and the body-size distance ratio parameter are parsed, converting the direction description into a three-dimensional unit direction vector. The body-size distance ratio parameter is converted into pixel distance values ​​based on a preset children's body shape parameter mapping table. This preset children's body shape parameter mapping table is based on children's growth curve parameters published by medical centers. It dynamically associates the body-size ratio conversion coefficient with the children's age segments and height and weight percentiles to achieve accurate pixel distance value adaptation. A scalar multiplication of the unit direction vector and pixel distance value generates a three-dimensional coordinate offset vector. The binary node connection matrix and the three-dimensional coordinate offset vector are combined and encapsulated into structured topological parameter data, i.e., the preprocessed TCM acupoint mapping rules. The acupoint names in the acupoint effect data are identified using international standard acupoint coding to generate unique acupoint identification codes; the massage intensity and frequency parameters in the acupoint effect text data are converted into a standardized set of action instructions; and the unique acupoint identification codes and standardized action instruction sets are encapsulated in JSON-LD format.

[0028] S2. Input the video stream image data into the pre-trained child pose estimation model. The child pose estimation model performs coordinate decoding on the video stream image data and outputs the joint coordinates.

[0029] The pre-trained child pose estimation model refers to collecting a child human body dataset, labeling the child-specific key points in the child human body dataset, inputting the labeled child human body dataset into the HRNet-W16 network architecture for transfer learning, and generating a pre-trained child pose estimation model. Specifically, the children's anatomy dataset contains dynamic images of children of multiple age groups, covering various postures such as standing, walking, and sitting, ensuring the diversity of the dataset to accommodate children's body shapes and movement characteristics. Each image frame in the dataset is labeled with child-specific joint points, including key anatomical locations such as the shoulder, elbow, wrist, hip, knee, and ankle. The labeling process was completed by TCM and anatomists based on children's growth and development characteristics, ensuring that the joint point locations are correlated with TCM acupoint mapping rules. The labeled children's anatomy dataset is then input into the HRNet-W16 network architecture, a high-resolution feature extraction network. The HRNet-W16 network architecture maintains the high resolution of feature maps through parallel multi-resolution sub-networks, making it suitable for high-precision detection of children's joints. The architecture employs a transfer learning strategy, pre-training on publicly available large-scale adult pose estimation datasets (such as the COCO dataset) to acquire general human feature extraction capabilities. It is then fine-tuned on labeled children's human datasets to optimize network weights to adapt to the characteristics of children's body shape and dynamic poses. The transfer learning process iteratively optimizes network parameters through gradient descent algorithm, using labeled children-specific joints as supervision signals to adjust the weights of convolutional and deconvolutional layers in the HRNet-W16 network architecture, generating a pre-trained children's pose estimation model.

[0030] The video stream image data is input into a pre-trained child pose estimation model. The child pose estimation model extracts features from the video stream image data through a convolutional neural network layer to generate multi-scale feature maps. Specifically, the pre-trained child pose estimation model is based on the HRNet-W16 network architecture and is trained on a child human body dataset through transfer learning, specifically optimized for children's joint features. The pre-trained child pose estimation model extracts features from each frame of standardized video stream image data through a multi-layer convolutional neural network. The convolutional neural network layer contains multiple convolutional kernels, capturing low-level features (such as edges and textures) and high-level features (such as human body contours and joint positions) in the image layer by layer. The convolutional neural network layer adopts a multi-scale processing method, extracting features in parallel through convolutional kernels of different sizes to generate multi-scale feature maps with multiple resolutions, preserving detailed information from low resolution to high resolution. The multi-scale feature maps contain the spatial distribution information of the child's body joints, such as feature representations of key anatomical regions like the shoulders, elbows, and knees, providing basic data for subsequent joint probability distribution generation. The generation process of multi-scale feature maps ensures the robustness of feature extraction and can adapt to the diversity of children's body shape changes and dynamic postures. Deconvolution and probabilistic activation are performed on the multi-scale feature map to output the probability distribution of key points; The multi-scale feature maps are input into the deconvolution process of the pre-trained child pose estimation model. The deconvolution process upsamples the multi-scale feature maps to recover spatial details and generate spatial probability distributions of joints. The deconvolution process is implemented through transposed convolution, using a convolution kernel size matched to the convolutional neural network layers, and upsampling the multi-scale feature maps layer by layer, fusing the multi-scale feature maps of different resolutions into a unified high-resolution feature representation. The high-resolution feature representation is processed by a probabilistic activation function, which uses the sigmoid function to map each pixel value of the multi-scale feature map to a probability value between 0 and 1, generating a joint probability distribution. The joint probability distribution represents the probability that each pixel belongs to a specific joint (such as the shoulder and knee), specifically in the form of a heatmap, with each heatmap corresponding to a joint. High-value regions in the heatmap indicate the possible locations of joints. The joint probability distribution preserves the spatial position information of the child's body in dynamic images, can adapt to changes in posture and body shape, and the output joint probability distribution serves as the input data for the next step of peak coordinate analysis.

[0031] Perform peak coordinate analysis on the probability distribution of key points and output the coordinates of the key points; Specifically, the probability distribution of joint points is generated by deconvolution and probabilistic activation processing, and is represented as a heatmap of the probability distribution of each joint point on the child's body. Each heatmap corresponds to one joint point, and the high-value areas in the heatmap reflect the possible locations of the joint points. Peak coordinate analysis is performed on the probability distribution of joint points to extract the location coordinates of the joint points from each heatmap. Peak coordinate analysis applies a non-maximum suppression algorithm to each heatmap. The non-maximum suppression algorithm traverses the pixel values ​​of the heatmap, identifies local maximum points, suppresses non-maximum regions, and retains the highest probability location of each joint point. Two-dimensional pixel coordinates are extracted from the highest probability location of each heatmap to generate joint point coordinates. The joint point coordinates are represented as two-dimensional vectors, corresponding to the pixel positions in the standardized video stream image data. Each joint point coordinate corresponds one-to-one with the joint point name, forming a joint point coordinate set. The joint point coordinate set contains the location information of key anatomical points of the child's body in dynamic images, such as the shoulders, elbows, and knees, which can reflect the child's dynamic posture and provide basic data for subsequent acupoint location.

[0032] S3. Extract the motion vector field between consecutive frames of video stream image data using the sparse optical flow algorithm, identify the trajectory of children's limb movements, generate a spatiotemporal displacement matrix, compensate and fuse the joint coordinates with the spatiotemporal displacement matrix, and output an optimized set of joint coordinates.

[0033] Feature point detection is performed on gradient regions in video stream image data, gradient coordinate sets are output, and pixel displacement vector statistics are performed on the gradient coordinate sets of adjacent frames to generate motion vector fields. Specifically, for each frame of the video stream image data, a sparse optical flow algorithm is applied for feature point detection. The sparse optical flow algorithm calculates the gradient region of each frame, and the gradient region is used to extract the intensity changes of pixels in each frame using the Sobel operator to identify high contrast edges and textures, such as the outline of a child's limb or high contrast regions near joints. The Sobel operator calculates the gradient in the horizontal and vertical directions for each frame, generating a gradient magnitude map. High-value regions in the gradient magnitude map represent feature point locations. The feature point detection process extracts all pixels with significant gradients, generating a gradient coordinate set. The gradient coordinate set contains the two-dimensional pixel coordinates of key regions of the child's limb in each frame. The gradient coordinate sets of adjacent frames are matched, and the KLT (Kanade-Lucas-Tomasi) tracking algorithm is used to obtain the pixel displacement of feature points between consecutive frames, determining the displacement direction and distance of each feature point, and generating a pixel displacement vector. The pixel displacement vector is represented in two-dimensional vector form, including displacement direction and displacement amount. The pixel displacement vectors of all feature points are summarized to generate a motion vector field, which describes the dynamic changes of the child's limb between consecutive frames. The formula for the gradient in the horizontal direction is: ; The formula for the gradient in the vertical direction is: ; in, This represents the horizontal coordinate identifier of the current pixel in each frame of the image. This represents the y-coordinate identifier of the current pixel in each frame of the image. This represents the row index in the convolution kernel. This represents the column index in the convolution kernel. Represents each frame of an image as pixels. Add the row and column indices of the neighboring pixel values ​​to the center. This represents the horizontal Sobel convolution kernel weights. This represents the horizontal gradient of each frame of the image. This represents the weights of the Sobel convolution kernel in the vertical direction. This represents the vertical gradient of each frame of the image; The formula for gradient magnitude is: ; The gradient direction formula is: ; in, This indicates the pixel position of each frame of the image. gradient magnitude at that point This indicates the pixel position of each frame of the image. The gradient direction.

[0034] The motion direction and speed of the child's limbs in the motion vector field are correlated and smoothed using a temporal clustering algorithm, and the motion trajectory of the child's limbs is output. Specifically, the motion vector field is input into a temporal clustering algorithm, which employs the DBSCAN (Density-Based Spatial Clustering) method. This method clusters the motion vectors based on their direction and velocity, using these as feature vectors. Based on the Euclidean distance and density state between feature points, pixel displacement vectors with similar directions and velocities are grouped into the same trajectory cluster. The DBSCAN algorithm then clusters these pixel displacement vectors into the same trajectory, associating them based on spatial proximity and temporal continuity of the feature points to generate preliminary information about children's limb movements. The initial trajectory of a child's limb movement is represented by a sequence of pixel displacement vectors over time, containing the positional changes of each feature point in consecutive frames. To improve the smoothness and continuity of the initial trajectory, a Kalman filter algorithm is applied. The Kalman filter algorithm estimates the true motion state of the feature points through prediction and updating, smooths abrupt changes in motion direction and speed, and eliminates trajectory jitter caused by changes in lighting or occlusion. After smoothing, the child's limb movement trajectory is output, which contains the stable displacement path of each feature point over time, reflecting the dynamic movement patterns of the child's limbs, such as arm swinging or leg movement. The trajectory of children's limb movements is encoded according to the time sequence of video stream image data to generate a spatiotemporal displacement matrix; The movement trajectories of children's limbs are encoded according to the time series of standardized video stream image data. The encoding process uses the timestamp of each frame as a reference, and decomposes the displacement path of each feature point into displacement vectors in the time and spatial dimensions. The time dimension is based on a frame rate of 30 frames per second of standardized video stream image data, and the spatial dimension is based on a two-dimensional coordinate system with a resolution of 512×512 pixels. For the displacement path of each feature point, the pixel displacement vector between consecutive frames is extracted. The displacement vector is represented in the form of (x, y) and includes the displacement in the horizontal and vertical directions. The displacement vectors of all feature points are arranged in the time series to construct a two-dimensional matrix. The rows of the matrix represent time frames, and the columns represent the displacement vectors of feature points, generating a spatiotemporal displacement matrix. Each row of the spatiotemporal displacement matrix corresponds to the displacement of a feature point in one frame of image, and each column corresponds to the displacement change of a feature point in the time series. The matrix elements are two-dimensional pixel displacement vectors, reflecting the spatiotemporal motion characteristics of children's limbs in dynamic images.

[0035] The coordinates of the joints are matched point by point with the spatiotemporal displacement matrix to generate joint-displacement data pairs. Specifically, the point-to-point matching process is based on spatial proximity and temporal consistency. For each key point coordinate, feature points within the same time frame in the spatiotemporal displacement matrix are searched, the Euclidean distance between the key point coordinate and the feature point coordinate is obtained, and the feature point with the closest Euclidean distance is selected as the matching object. The coordinates of the successfully matched key point are combined with the pixel displacement vector of the corresponding feature point to form a key point-displacement data pair. The key point-displacement data pair contains the key point name, pixel coordinates, and the corresponding pixel displacement vector. Based on the motion vector data in the joint-displacement data pair, the joint coordinates are corrected for displacement, and the compensated coordinates are output. The displacement correction process superimposes pixel displacement vectors onto joint coordinates to output compensated coordinates. The compensated coordinates are represented in two-dimensional vector form. During the process of superimposing pixel displacement vectors onto joint coordinates, a consistent resolution benchmark must be maintained. The corrected compensated coordinates reflect the true position of the joints in dynamic motion, correcting static coordinate offsets caused by children's limb movements, such as deviations caused by arm swings or leg movements. The compensated coordinates retain the joint name and two-dimensional pixel coordinate format, consistent with the original joint coordinates, facilitating subsequent processing. The compensation coordinates and the joint coordinates are weighted, averaged, and encapsulated to output an optimized set of joint coordinates. Specifically, the weighted averaging process is applied point-by-point to each joint, ensuring that the new coordinates comprehensively consider both the accuracy of static detection and the compensation effect of dynamic motion. The weighted averaged coordinates are represented in two-dimensional vector form, retaining the format of joint names and pixel resolution. To ensure the integrity and consistency of the results, boundary checks are performed on the new weighted averaged coordinates to prevent them from exceeding the pixel resolution boundaries. All weighted averaged coordinates are sorted by joint name and encapsulated into an optimized joint coordinate set. The optimized joint coordinate set contains the corrected pixel coordinates of all key anatomical points of the child's body, reflecting their true positions in the dynamic images. The optimized joint coordinate set provides high-precision input data for subsequent geometrically constrained initial acupoint localization.

[0036] It should be noted that by extracting the motion vector field of video stream image data through the sparse optical flow algorithm, and combining it with temporal clustering and Kalman filtering to generate a spatiotemporal displacement matrix, dynamic tracking of children's limb movement trajectories and compensation fusion of joint coordinates are achieved, outputting an optimized set of joint coordinates, thus improving the accuracy and robustness of acupoint localization in dynamic scenes. Compared with traditional static joint detection schemes, continuous frame motion analysis effectively reduces the offset error caused by children's limb movements, and the combination of traditional Chinese medicine acupoint mapping rules provides high-precision input for subsequent acupoint localization, reducing the positioning deviation problems caused by insufficient dynamic tracking accuracy and differences in children's body shape in existing schemes. It innovatively achieves real-time adaptive calibration, which meets the actual needs of children's massage acupoint recognition.

[0037] S4. Based on the optimized set of joint coordinates, perform regular mapping on the target acupoints of children according to the rules of traditional Chinese medicine acupoint mapping, and output the coordinates of the massage acupoints.

[0038] Using the mapping rules of traditional Chinese medicine acupoints, the target acupoints are associated and filtered by optimizing the coordinate set of joint points, and the coordinates of anatomical landmarks are output. Specifically, the optimized set of joint coordinates is associated with and filtered using traditional Chinese medicine (TCM) acupoint mapping rules. This association and filtering process extracts a binary node connection matrix from the TCM acupoint mapping rules. The row index of the matrix represents the acupoint name, and the column index represents the anatomical landmark name. For example, the Zusanli acupoint is associated with the knee eye and ankle joints. For each acupoint name, the corresponding anatomical landmark name is found in the binary node connection matrix to determine the joints associated with the target acupoint. For example, the knee eye and ankle joints are selected as the anatomical landmarks for the Zusanli acupoint. The coordinates of joints matching the anatomical landmark names are extracted from the optimized set of joint coordinates to generate anatomical landmark coordinates. These coordinates are represented as two-dimensional vectors, containing the joint name and pixel coordinates, such as knee eye coordinates and ankle coordinates. Based on the relative position description of different acupoints in the TCM acupoint mapping rules, the offset direction and acupoint distance of the target acupoint relative to the coordinates of the anatomical landmark are determined, and the offset direction and acupoint distance are integrated to generate acupoint offset data; For each target acupoint, based on the relative position description in the TCM acupoint mapping rules, the offset direction and acupoint distance relative to the anatomical landmark coordinates are determined. The relative position description includes the spatial relationship between the acupoint and the anatomical landmark, such as the Zusanli acupoint being located below the knee eye and above the ankle. The offset direction is directly extracted from the unit direction vector of the TCM acupoint mapping rules, which is represented in two-dimensional vector form and describes the spatial movement direction of the target acupoint relative to the anatomical landmark. The acupoint distance is obtained from the pixel distance value of the TCM acupoint mapping rules, which is based on the children's body shape parameter mapping table. The offset direction and acupoint distance are integrated to generate acupoint offset data, which is represented in two-dimensional vector form.

[0039] The coordinates of anatomical landmarks and the offset data of acupoints are combined and verified against the child's body surface boundary to output the coordinates of massage acupoints. Specifically, the positional relationship combination process derives the acupoint coordinates for each target acupoint. This derivation involves superimposing anatomical landmark coordinates with acupoint offset data. For example, the massage acupoint coordinates for Zusanli (ST36) are obtained by superimposing the knee eye coordinates (x1, y1) with the acupoint offset data (0, -30), resulting in (x1, y1-30). The positional relationship combination is applied one by one to each target acupoint, generating massage acupoint coordinates represented as a two-dimensional vector, preserving the target acupoint name and fixed pixel resolution. To ensure the rationality of the massage acupoint coordinates, a child's body surface boundary verification is performed. This verification process checks whether the massage acupoint coordinates are within the valid range of the standardized video stream image data. The effective range is within the pixel resolution of the video stream image data, such as 512×512 pixel resolution. For massage acupoint coordinates that exceed the fixed pixel resolution boundary, a boundary clipping method is used to restrict the coordinate values ​​to the range of 0 to 512 pixels, such as setting negative values ​​to 0 and values ​​exceeding 512 to 512. The massage acupoint coordinates after boundary verification reflect the true pixel position of the target acupoint on the child's body surface, adapting to changes in body shape and posture in dynamic images. All massage acupoint coordinates are sorted according to the acupoint name and encapsulated into a massage acupoint coordinate set, which contains the pixel coordinates of all target acupoints, providing high-precision input data for subsequent acupoint topology optimization.

[0040] S5. Construct an acupoint topology map based on the coordinates of massage acupoints and the optimized set of joint point coordinates. Use a graph convolutional neural network to perform spatial relationship reasoning on the acupoint topology map and output a high-precision acupoint coordinate set.

[0041] The coordinates of massage acupoints and the optimized set of joint point coordinates are used as acupoint nodes, and node association edges between acupoint nodes are constructed according to the mapping rules of traditional Chinese medicine acupoints to generate an acupoint topology graph. Specifically, the coordinates of massage acupoints and the optimized set of joint point coordinates are used as acupoint nodes. These nodes are divided into two categories: one category consists of acupoint nodes corresponding to massage acupoint coordinates, such as the Zusanli acupoint node; the other category consists of acupoint joint points corresponding to the optimized set of joint point coordinates, such as the knee eye and ankle nodes. Each acupoint node contains a node name and two-dimensional pixel coordinates, based on a fixed pixel resolution. Node association edges are constructed between acupoint nodes based on traditional Chinese medicine acupoint mapping rules. These edges are represented as undirected edges, connecting related acupoint nodes and acupoint joint points, reflecting spatial proximity and anatomical relationships. For example, the edge between the Zusanli acupoint node and the knee eye joint point represents their proximity on the body surface. All acupoint nodes and their associated edges are combined to form an acupoint topological relationship graph. This graph is represented by a graph structure, where nodes store pixel coordinates and names, and edges store spatial association information, providing basic data for subsequent spatial relationship reasoning.

[0042] The acupoint topology map is input into a graph convolutional neural network, and the neighborhood features of each acupoint node are aggregated through a multi-layer graph attention mechanism to generate acupoint feature vectors. Specifically, the acupoint topology map is input into a graph convolutional neural network (GNN). The GNN, based on the GAT (Graph Attention Network) architecture, processes the acupoint topology map and aggregates node features through a multi-layer graph attention mechanism. The GNN initializes a feature vector for each acupoint node, generated from the node's pixel coordinates and represented as a two-dimensional vector. This initial feature vector is based on a fixed pixel resolution of standardized video stream image data; for example, the initial feature vector for the Zusanli acupoint node is (x, y). The multi-layer graph attention mechanism processes the acupoint topology map layer by layer. Each layer aggregates the feature information of neighboring nodes by acquiring the spatial relationships between the acupoint node and its neighbors. The aggregation process is completed through weighted summation. The weighted summation is based on the spatial proximity of the nodes' associated edges, prioritizing spatially close neighboring acupoint nodes. For example, the features of the knee eye and ankle joints are aggregated first for the Zusanli acupoint node. Each layer of the graph attention mechanism updates the initial feature vector of the acupoint node, generating richer feature representations, including the pixel coordinates of the acupoint node itself and the spatial relationship information of neighboring acupoint nodes. The multi-layer graph attention mechanism enhances the expressive power of the initial feature vector through continuous processing, generating acupoint feature vectors. The acupoint feature vectors are represented in the form of multi-dimensional vectors, containing the spatial location of each acupoint node and its association information with neighboring acupoint nodes.

[0043] Spatial calibration of acupoint coordinates is performed based on acupoint feature vectors to output a high-precision acupoint coordinate set. Specifically, the spatial calibration process processes acupoint feature vectors through a fully connected layer, mapping the multi-dimensional feature vectors back to a two-dimensional pixel coordinate space. The fully connected layer receives the acupoint feature vectors, processes each acupoint node one by one, and adjusts the pixel coordinates of the acupoint nodes through linear transformation and nonlinear activation functions to generate calibrated massage acupoint coordinates. The calibration process comprehensively considers the spatial relationship information in the acupoint feature vectors to correct deviations in massage acupoint coordinates caused by dynamic movement or body shape differences. For example, the coordinates of Zusanli acupoint are fine-tuned according to the spatial relationship between the knee joint and the ankle joint. The calibrated massage acupoint coordinates are represented in two-dimensional vector form, forming a high-precision acupoint coordinate set. The high-precision acupoint coordinate set contains the corrected pixel coordinates of all target acupoints, reflecting the precise acupoint positions on the child's body surface in dynamic images. The high-precision acupoint coordinate set provides high-precision input data for subsequent three-dimensional coordinate transformation and visualization output.

[0044] S6. Perform three-dimensional coordinate transformation on the high-precision acupoint coordinate set, and overlay the high-precision acupoint coordinate set after three-dimensional coordinate transformation and acupoint effect data onto the video stream image data to generate an acupoint visualization interface.

[0045] A three-dimensional coordinate transformation is performed on the high-precision acupoint coordinate set to generate a spatial acupoint coordinate set, and the spatial acupoint coordinate set is associated with the acupoint effect data to generate acupoint association data; Specifically, the 3D coordinate transformation process uses the intrinsic and extrinsic parameters of the RGB camera device to map the 2D pixel coordinates to the 3D world coordinate system. The intrinsic parameters of the RGB camera device include focal length and principal point coordinates, while the extrinsic parameters include the rotation and translation parameters of the RGB camera device relative to the child's body. The 3D coordinate transformation process projects the 2D pixel coordinates of the high-precision acupoint coordinate set onto the camera coordinate system to generate preliminary 3D coordinates. The preliminary 3D coordinates are then combined with a child's body shape parameter mapping table to adjust the scale of the 3D coordinates. The child's body shape parameter mapping table determines the body size ratio conversion factor based on age segments and height and weight percentiles. For example, the body size ratio of a 3-year-old child is converted to 10 millimeters. The adjusted 3D coordinates form a spatial acupoint coordinate set, which is represented in the form of a 3D vector, containing the target acupoint name and 3D coordinates, reflecting the actual spatial location of the acupoint on the child's body surface. The spatial acupoint coordinate set is associated with acupoint function data. The association process is carried out by matching the unique identifier of the acupoint name. The spatial acupoint coordinates are combined with the corresponding acupoint function classification, associated diseases and standardized action instruction set to generate acupoint association data. The acupoint association data includes acupoint name, three-dimensional coordinates and acupoint function information, and retains JSON-LD format to provide basic data for subsequent visualization overlay.

[0046] The acupoint-related data is overlaid onto the video stream image data, and the overlaid video stream image data is visually marked, rendered, and interactively added to generate an acupoint visualization interface. Specifically, the overlay process projects the three-dimensional coordinates in the acupoint-related data back to two-dimensional pixel coordinates. Using the aforementioned camera intrinsic and extrinsic parameters, the projection maps the three-dimensional coordinates to a 512×512 pixel resolution plane of the standardized video stream image data, generating two-dimensional acupoint coordinates corresponding to each frame. These two-dimensional acupoint coordinates are bound to the acupoint name and its function information and embedded into the corresponding frame of the video stream image data. Visual marker rendering is then applied to the overlaid video image data, drawing pulsating halo markers at the two-dimensional acupoint coordinate positions. These pulsating halo markers are represented by circular shapes, with their centers located at the two-dimensional acupoint coordinates. The colors used are high-contrast hues (such as green), and periodic changes in transparency simulate a pulsating effect, enhancing visual recognizability. The visual marker rendering process utilizes image recognition methods to ensure that the pulsating halo markers accurately align with the acupoint positions on the child's body surface, adapting to posture in dynamic images. The changes include: adding a transparent touch-responsive layer to the pulsating halo markers, covering the pulsating halo marker area, and associating it with pre-recorded audio explanations of children's acupoints. The audio content includes acupoint names, functional classifications, and massage instructions, such as the audio explanation of the effects of Zusanli acupoint on regulating the spleen and stomach and gentle massage methods. The transparent touch-responsive layer triggers audio playback by detecting user touch input, and the audio is stored in JSON-LD format in the acupoint association data, corresponding one-to-one with the acupoint names. The interaction addition process incorporates posture perception, tracking changes in children's body posture and dynamically adjusting the positions of the pulsating halo markers and touch-responsive layer to ensure synchronization with dynamic images. The visual marker rendering and the video stream image data after the interaction addition generate an acupoint visualization interface, presented as a video stream at 30 frames per second, including pulsating halo markers and an interactive touch-responsive layer, providing intuitive massage guidance for children and parents.

[0047] This embodiment also provides an artificial intelligence-based acupoint recognition system for children's massage, including: The data acquisition module collects children's massage data and performs preprocessing; the children's massage data includes video stream image data, traditional Chinese medicine acupoint mapping rules, and acupoint effect data; The joint module is used to input video stream image data into a pre-trained child pose estimation model. The child pose estimation model performs coordinate decoding on the video stream image data and outputs the coordinates of the joint points. The optimization module is used to extract the motion vector field between consecutive frames of video stream image data through the sparse optical flow algorithm, identify the trajectory of children's limb movements, generate a spatiotemporal displacement matrix, compensate and fuse the joint coordinates with the spatiotemporal displacement matrix, and output an optimized set of joint coordinates. The acupoint module is used to perform rule mapping on target acupoints for children based on an optimized set of joint coordinates and according to the rules of traditional Chinese medicine acupoint mapping, and output the coordinates of the massage acupoints. The graph module is used to construct an acupoint topology graph based on the coordinates of massage acupoints and an optimized set of joint point coordinates. It uses a graph convolutional neural network to perform spatial relationship reasoning on the acupoint topology graph and outputs a high-precision set of acupoint coordinates. The interface module is used to perform three-dimensional coordinate transformation on the high-precision acupoint coordinate set, and then overlay the high-precision acupoint coordinate set and acupoint effect data into the video stream image data to generate an acupoint visualization interface.

[0048] This embodiment also provides a computer device applicable to the artificial intelligence-based children's massage acupoint recognition method, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to realize the artificial intelligence-based children's massage acupoint recognition method as proposed in the above embodiment.

[0049] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0050] This embodiment also provides a storage medium storing a computer program, which, when executed by a processor, implements the artificial intelligence-based acupoint recognition method for children's massage as proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0051] In summary, this invention achieves real-time, accurate positioning and adaptive calibration of acupoints for children's massage through a dual mechanism of dynamic motion compensation and traditional Chinese medicine topology optimization. A sparse optical flow algorithm is used to extract the motion vector field and compensate for it with joint coordinates, outputting an optimized set of joint coordinates. This dynamically tracks the trajectory of children's limbs and eliminates posture offset errors in real time, ensuring continuous accuracy of acupoint positioning in dynamic scenarios. A graph convolutional neural network is used to perform spatial relationship reasoning on the acupoint topology map, modeling anatomical constraints and spatial association rules between acupoints. Based on neighborhood feature aggregation, the spatial coordinate distribution of acupoints is adaptively calibrated, reducing acupoint mapping deviations caused by differences in children's growth and improving positioning robustness under complex body shapes.

[0052] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. An artificial intelligence-based child massage acupoint identification method, characterized in that: The application relates to a child massage data processing method and device. Collect child massage data and perform preprocessing; the child massage data comprises video stream image data, Chinese medicine acupoint mapping rules and acupoint action data; Input the video stream image data into a pre-trained child posture estimation model, the child posture estimation model performs coordinate decoding on the video stream image data, and outputs joint point coordinates; Extract a motion vector field between continuous frames of the video stream image data through a sparse optical flow algorithm, identify a child limb motion trajectory, generate a space-time displacement matrix, compensate and fuse the joint point coordinates and the space-time displacement matrix, and output an optimized joint point coordinate set; Based on the optimized joint point coordinate set, perform rule mapping on a child target acupoint according to the Chinese medicine acupoint mapping rules, and output massage acupoint coordinates; Construct an acupoint topological relationship graph according to the massage acupoint coordinates and the optimized joint point coordinate set, perform spatial relationship reasoning on the acupoint topological relationship graph by adopting a graph convolutional neural network, and output a high-precision acupoint coordinate set; Perform three-dimensional coordinate conversion on the high-precision acupoint coordinate set, superimpose the high-precision acupoint coordinate set after three-dimensional coordinate conversion and the acupoint action data on the video stream image data, and generate an acupoint visualization interface. 2.The artificial intelligence-based child massage acupoint recognition method of claim 1, wherein: The specific steps are as follows: Input the video stream image data into a pre-trained child posture estimation model, the child posture estimation model performs coordinate decoding on the video stream image data, and outputs joint point coordinates, and the specific steps are as follows: Input the video stream image data into a pre-trained child posture estimation model, the child posture estimation model performs coordinate decoding on the video stream image data, and outputs joint point coordinates, and the specific steps are as follows: The pre-trained child posture estimation model is generated by collecting a child human body data set, labeling special joint points of children in the child human body data set, inputting the labeled child human body data set into an HRNet-W16 network architecture, and performing migration learning. 3.The artificial intelligence-based child massage acupoint recognition method of claim 2, wherein: The specific steps are as follows: 4.The artificial intelligence-based child massage acupoint recognition method of claim 3, characterized in that: Detect feature points in gradient regions in the video stream image data, output gradient coordinate sets, and perform pixel displacement vector statistics on gradient coordinate sets of adjacent frames to generate a motion vector field; Perform trajectory association and smoothing processing on the motion direction and motion speed of the child limb in the motion vector field through a time series clustering algorithm, and output a child limb motion trajectory; Encode the child limb motion trajectory according to the time sequence of the video stream image data to generate a space-time displacement matrix. The specific steps are as follows: 5.The artificial intelligence-based child massage acupoint recognition method of claim 4, characterized in that: Match the joint point coordinates and the space-time displacement matrix point by point to generate joint point-displacement data pairs; Displacement correct the joint point coordinates based on the motion vector data in the joint point-displacement data pairs, and output compensation coordinates; Perform weighted average and packaging on the compensation coordinates and the joint point coordinates, and output an optimized joint point coordinate set. ​ 6.The artificial intelligence-based child massage acupoint recognition method of claim 5, wherein: The child target acupoint is mapped according to the TCM acupoint mapping rule based on the optimized key joint coordinate set, and massage acupoint coordinates are output, and the specific steps are, The target acupoint is associated and screened from the optimized key joint coordinate set by using the TCM acupoint mapping rule, and anatomical landmark point coordinates are output; Based on the relative position description of different acupoints in the TCM acupoint mapping rule, the offset direction and acupoint distance of the target acupoint relative to the anatomical landmark point coordinates are determined, and the offset direction and acupoint distance are integrated to generate acupoint offset data; The anatomical landmark point coordinates and the acupoint offset data are combined in terms of position relationship and checked against the child body surface boundary, and massage acupoint coordinates are output.

7. The artificial intelligence-based child massage acupoint recognition method of claim 6, wherein: The massage acupoint coordinates and the optimized key joint coordinate set are used to construct an acupoint topological relationship graph, a graph convolutional neural network is used to reason the spatial relationship of the acupoint topological relationship graph, and a high-precision acupoint coordinate set is output, and the specific steps are, The massage acupoint coordinates and the optimized key joint coordinate set are used as acupoint nodes, and node association edges between the acupoint nodes are constructed according to the TCM acupoint mapping rule to generate an acupoint topological relationship graph; The acupoint topological relationship graph is input into the graph convolutional neural network, the neighborhood features of each acupoint node are aggregated through a multi-layer graph attention mechanism, and an acupoint feature vector is generated; Based on the acupoint feature vector, the massage acupoint coordinates are spatially calibrated, and a high-precision acupoint coordinate set is output. 8.The artificial intelligence-based child massage acupoint recognition method of claim 7, wherein: The high-precision acupoint coordinate set is converted into three-dimensional coordinates, and the high-precision acupoint coordinate set after three-dimensional coordinate conversion and acupoint action data are superimposed into video stream image data to generate an acupoint visualization interface, and the specific steps are, The high-precision acupoint coordinate set is converted into three-dimensional coordinates to generate a spatial acupoint coordinate set, and the spatial acupoint coordinate set is associated with the acupoint action data to generate acupoint association data; The acupoint association data is superimposed into the video stream image data, and the superimposed video stream image data is rendered and interactively added with visual markers to generate an acupoint visualization interface. 9.The artificial intelligence-based child massage acupoint recognition method of claim 8, wherein: The visual marker rendering and interactive addition refers to drawing a pulsating light ring marker at the coordinate position of the acupoint association data, and adding a transparent touch response layer in the pulsating light ring marker, the transparent touch response layer is associated with a pre-recorded child acupoint explanation audio to generate an acupoint visualization interface.

10. An artificial intelligence-based child massage acupoint recognition system based on any one of the artificial intelligence-based child massage acupoint recognition methods of claims 1-9. It includes, The acquisition module acquires child massage data and performs preprocessing; The child massage data includes video stream image data, TCM acupoint mapping rules and acupoint action data; The joint module is used to input the video stream image data into a pre-trained child pose estimation model, and the child pose estimation model decodes the coordinates of the video stream image data to output key joint coordinates; The optimization module is used to extract the motion vector field between consecutive frames of the video stream image data by using a sparse optical flow algorithm, identify the child limb motion trajectory, generate a space-time displacement matrix, compensate and fuse the key joint coordinates and the space-time displacement matrix, and output an optimized key joint coordinate set; The acupoint module is used to map the child target acupoint according to the TCM acupoint mapping rule based on the optimized key joint coordinate set, and output massage acupoint coordinates; The atlas module is configured to construct an acupoint topological relationship atlas according to the massage acupoint coordinates and the optimized node coordinate set, perform spatial relationship reasoning on the acupoint topological relationship atlas by using a graph convolutional neural network, and output a high-precision acupoint coordinate set; The interface module is configured to perform three-dimensional coordinate conversion on the high-precision acupoint coordinate set, superimpose the high-precision acupoint coordinate set after three-dimensional coordinate conversion and acupoint action data into video stream image data, and generate an acupoint visualization interface.

Images