CN117974740A - Acupoint positioning method and robot based on aggregation type window self-attention mechanism - Google Patents

Abstract

The invention discloses an acupoint positioning method based on an aggregated window self-attention mechanism. First, a human back image is obtained and preprocessed. The preprocessed back image is input into a multi-scale feature extraction network to obtain a back feature map. The multi-scale feature extraction network includes several feature map extraction modules. The preprocessed back image is processed by several feature map extraction modules in turn. Each feature map extraction module includes an aggregated window self-attention learning layer. The back feature map is input into a back feature key point detection network to perform back feature key point detection to obtain back feature key points. The specific coordinates of the back acupoints are obtained through a back acupoint positioning formula to achieve back acupoint positioning. In obtaining the features of the back feature key points of the human body, the aggregated window self-attention learning method is used for the position of the back where the key points are located, so that the key point positions of the back of human bodies of different body shapes can be determined more accurately and quickly.

Description

Acupoint localization method and robot based on aggregated window self-attention mechanism

Technical Field

The present invention relates to a physiotherapy robot, and in particular to an acupoint positioning method and a robot based on a clustered window self-attention mechanism.

Background technique

Acupoints are an important part of medicine. Massaging acupoints can prevent and cure diseases and improve health. Physicians usually use their long-term massage experience to select and press acupoints, but this method requires physicians to have long-term experience in selecting acupoints, and there are problems such as a small number of clinical doctors who learn Chinese massage and poor medical skills. In recent years, with the development of computer vision technology and robots, some researchers have proposed methods based on machine vision and image processing to locate acupoints, and applied them to robots to realize intelligent acupoint selection and pressing.

However, most of today's massage robots based on visual positioning often have very high requirements for the working environment, require a single massage background, are not universal for complex backgrounds, and have low positioning accuracy. They are still far from professional masseurs in terms of intelligence.

Summary of the invention

Purpose of the invention: In view of the above shortcomings, the present invention provides an acupoint positioning method and robot based on a clustered window self-attention mechanism with high positioning accuracy.

Technical solution: To solve the above problems, the present invention adopts an acupoint positioning method based on an aggregated window self-attention mechanism, comprising the following steps:

(1) Acquire a human back image and perform preprocessing to obtain a preprocessed back image;

(2) inputting the preprocessed back image into a multi-scale feature extraction network to obtain a back feature map; the multi-scale feature extraction network includes a plurality of feature map extraction modules, the preprocessed back image is sequentially processed by the plurality of feature map extraction modules, each feature map extraction module includes an aggregated window self-attention learning layer, and the aggregated window self-attention learning layer is used to strengthen feature learning of pixels containing feature key points in the extracted feature map;

(3) Inputting the back feature map into the back feature key point detection network to detect the back feature key points and obtain the back feature key points;

(4) Based on the key points of the back features, the specific coordinates of the back acupoints are obtained through the back acupoint positioning formula to achieve back acupoint positioning.

Further, the preprocessing of the human back image includes:

Obtaining depth information of the human back image;

The depth information is fused with the human back image through the depth information fusion formula to obtain a visual depth information map;

The binary image of the human back is used as a mask to compensate for the lack of visual depth information map and obtain the preprocessed back image.

Further, the obtaining of the depth information of the human back image specifically includes: obtaining a disparity map of the human back image, converting the disparity map into a depth map by a depth conversion formula to obtain the depth information, and the depth information calculation formula is:

;

in, is the depth information, /> is parallax, /> is the baseline length, /> is the focal length, /> 、/> They are the horizontal coordinates of the same positions in the disparity map corresponding to the left and right cameras of the depth binocular camera.

Furthermore, the depth information fusion formula is:

;

in, is the depth information of the i-th pixel in the human back image, /> is the horizontal coordinate of the ith pixel in the human back image, /> is the ordinate of the i-th pixel in the human back image, /> is the horizontal coordinate of the intersection of the camera optical axis and the imaging plane, /> is the ordinate of the intersection of the camera optical axis and the imaging plane, /> 、/> 、/> is the fusion point coordinate of the i-th pixel in the visual depth information map, /> 、/> They are respectively/> Axis direction and /> The length of a single pixel in the axis direction.

Further, the preprocessed back image is composed of a human back image Binarized image of and visual depth information map/> Use binary degree number/> The random fusion is obtained, and the fusion formula is:

;

in, is the regression reward model,/> is an infinite non-repeating decimal, approximately equal to ,/> For the degree of reward.

Furthermore, the multi-scale feature extraction network includes a primary feature map extraction module, a secondary feature map extraction module, a tertiary feature map extraction module, and a quaternary feature map extraction module; the preprocessed back image is input into the primary feature map extraction module, and a primary feature map is output; the primary feature map extraction module includes a patch partition transformation layer and a first aggregated window self-attention learning layer; the patch partition transformation layer includes patch cropping and depth compression of the preprocessed back image;

The primary feature map is input into a secondary feature map extraction module, and a secondary feature map is output, wherein the secondary feature map extraction module includes a first patch merging layer and a second aggregated window self-attention learning layer;

The secondary feature map is input into a third-level feature map extraction module, and a third-level feature map is output. The third-level feature map extraction module includes a second patch merging layer and a third aggregated window self-attention learning layer;

The three-level feature map is input into a four-level feature map extraction module, and a four-level feature map, i.e., a back feature map, is output. The four-level feature map extraction module includes a third patch merging layer and a fourth aggregated window self-attention learning layer; the first patch merging layer, the second patch merging layer, and the third patch merging layer are used to perform non-convolutional downsampling on the feature map.

Furthermore, the first aggregated window self-attention learning layer, the second aggregated window self-attention learning layer, the third aggregated window self-attention learning layer, and the fourth aggregated window self-attention learning layer all include the following steps:

Divide the input feature map into windows to obtain several windows;

Perform self-attention learning on each window separately;

Shift each window after self-attention learning to the four corners to form an aggregate window on the corners of each window;

Aggregate self-attention learning is performed inside the aggregation window, and the enhanced feature map is output.

Furthermore, the self-attention learning is to use the vector corresponding to the information of each pixel in the single-channel feature map as the input vector, generate the corresponding information matrix through the weight matrix, and input the information matrix into the self-attention learning function;

The formula for the aggregated self-attention learning process is:

;

in, The top left corner of the aggregation window/> Row, No./> The feature matrix of the column-aggregated window pixels; /> and is the feature matrix of pixels aggregated inside the window; /> is the feature matrix/> The back feature quantity in ,/> is the feature matrix/> The back feature quantity in ;

When it appears When , the aggregation window only has back region features and no background region features, then the aggregation window is screened out;

When the characteristic matrix The back feature quantity of is given by/> Growth to/> And when it is the only maximum value among the feature quantities of the four aggregation windows, it is determined that the aggregation window is the area in the four aggregation windows that is most likely to have back feature key points.

Furthermore, the back feature key point detection network includes a global pooling layer and a classifier; the global pooling layer performs global average pooling on the back feature map, and the back feature map is compressed into a feature vector of a fixed length. The classifier inputs the feature vector into the fully connected layer to obtain the predicted labels and coordinates of the back feature key points.

The present invention also adopts a robot applying the above-mentioned acupoint positioning method, comprising an image acquisition module, a processing module, a control module, and an end effector;

The image acquisition module is used to acquire the image of the back of the human body;

The processing module is used to preprocess the acquired human back image to obtain a preprocessed back image; input the preprocessed back image into a multi-scale feature extraction network to obtain a back feature map; input the back feature map into a back feature key point detection network to perform back feature key point detection to obtain back feature key points; according to the back feature key points, the specific coordinates of the back acupoints are obtained by the back acupoint positioning formula to realize back acupoint positioning;

The control module is used to control the end effector to perform kneading massage on the located acupuncture points on the back.

Beneficial effect: Compared with the prior art, the significant advantage of the present invention is that in acquiring the key features of the back of the human body, the position of the key points on the back where the key points are located is determined more accurately and quickly by using a clustered window self-attention learning method. The back acupuncture point positioning formula realizes automatic acupuncture point finding to solve the problems that most existing back acupuncture point positioning methods need to re-find acupuncture points for people of different body shapes, do not automatically locate acupuncture points, and have low acupuncture point positioning accuracy and large workload. In detecting the key features of the back of the human body, the binary image of the original RGB image is used as a mask to compensate for the lack of the depth image on the contour of the back of the human body, which can better divide the back area of the human body and facilitate the detection of the key features of the back.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG1 is a schematic diagram of the workflow of the acupoint locating method of the present invention.

FIG. 2 is an image of a human back captured by a camera in the present invention.

FIG. 3 is a visual depth information diagram obtained by visualizing depth information on a human back image in the present invention.

FIG. 4 is a depth compensation map obtained after contour compensation of the visible depth information map in the present invention.

FIG. 5 is a result of detecting key feature points on the back of a human body using the prior art without aggregated window self-attention learning.

FIG6 is a result of detecting key feature points of the back of a human body obtained by using the present invention through aggregated window self-attention learning.

Detailed ways

In this embodiment, the RealSense D435i deep binocular camera collects RGB images of the human back, processes the images through the software of the host computer, displays the processing results on the LCD screen, and inputs the processing results into the robot processor for kneading massage. This method can be applied to the fields of human acupoint positioning and robot autonomous massage.

As shown in FIG1 , in this embodiment, an acupoint positioning method based on an aggregated window self-attention mechanism includes the following steps:

(1) The robot uses a deep binocular camera to collect RGB images of the human back and perform preprocessing;

(2) The preprocessed back image is input into the multi-scale feature extraction network, and passes through the first-level feature map extraction module, the second-level feature map extraction module, the third-level feature map extraction module, and the fourth-level feature map extraction module in the network in sequence, and finally outputs a fourth-level feature map;

(3) The four-level feature map is input into the back feature key point detection network to detect the back feature key points of the preprocessed back image, and the specific coordinates of the back acupoints are obtained through the back acupoint positioning formula.

The present invention will be further described below in conjunction with the accompanying drawings and specific embodiments.

(1) The collected RGB image of the human back is shown in FIG2 . The human back RGB image, original human back RGB image, and original RGB image mentioned in all the following steps refer to the image in FIG2 .

Preprocessing of the RGB image of the human back: In order to better segment the human back area from the background area, the depth information in the original RGB image is extracted and fused with the original image through the depth information fusion formula to obtain a visual depth information map. The binary image of the original RGB image is used as a mask to compensate for the lack of the depth image on the human back contour to obtain a compensated depth map. Specifically, the following steps are performed:

(2.1) Based on the original RGB image collected by the deep binocular camera, the corresponding disparity map is obtained, and the disparity map is converted into a depth map through the depth conversion formula to obtain the depth information. The depth calculation formula is:

;

in, is the depth information, /> is parallax, /> is the baseline length, /> is the focal length, /> 、/> They are the horizontal coordinates of the same positions in the disparity map corresponding to the left and right cameras of the depth binocular camera.

The depth information of the depth map is fused into the original RGB image to obtain a visual depth information map, as shown in Figure 3. The calculation formula is:

;

in, is the depth information of the i-th pixel in the original human back RGB image, /> is the horizontal coordinate of the i-th pixel in the original human back RGB image, /> is the ordinate of the i-th pixel in the original human back RGB image, /> is the horizontal coordinate of the intersection of the camera optical axis and the imaging plane, /> is the ordinate of the intersection of the camera optical axis and the imaging plane, /> 、/> 、/> is the fusion point coordinate of the i-th pixel in the visual depth information map, /> 、/> They are respectively/> Axis direction and /> The length of a single pixel in the axis direction.

(2.3) Based on the depth map, the original RGB image is first binarized to obtain a binary image This image is used as a mask to compensate for the lack of depth image on the back contour of the human body. Specifically, the depth image is compensated It is a binary image/> and visual depth information map/> Use any binary degree number/> Random fusion gives:

;

in, ,/> To be compensated for the Row, No./> Column pixels, /> ;

For the contour compensation process, disturbance occurs, that is, the binary image With visual depth information map /> In the case of misalignment during compensation, use /> The regression reward model is refined, and the refinement is done by/> Specified /> Area. Through the above operations, /> The regression reward model can focus on correcting the compensated depth map/> The depth value in each region can be completely retained after the disturbance is removed, and the boundary of the binary mask will not be damaged due to the removal of the disturbance. As shown in Figure 4, the depth compensation map is obtained after the visual depth information map is compensated.

(3) The depth compensation map obtained after compensation is input into the multi-scale feature extraction network, and passes through the first-level feature map extraction module, the second-level feature map extraction module, the third-level feature map extraction module, and the fourth-level feature map extraction module in the network in turn to finally obtain the fourth-level feature map including:

(3.1) First-level feature map extraction module: The preprocessed back image is sequentially input into the patch partitioning transformation layer and the first aggregated window self-attention learning layer, and the first-level feature map is output;

The patch partition transformation layer includes:

The preprocessed back image is first patch-cropped, that is, every 4×4 adjacent pixels in the preprocessed image are divided into a patch, and flattened in the color channel direction. Since each patch has 16 pixels, the input is an RGB three-channel image, and each pixel has three values of R, G, and B, so the flattened size is 16×3=48, and the image size is given by became , where /> is the height of the preprocessed image, /> is the width of the preprocessed image, and 48 is the depth of the preprocessed image; then a linear transformation is performed on the channel data of each pixel to compress the depth of the preprocessed image from 48 to /> , that is, the size of the preprocessed image is then determined by/> Became/> ,/> The depth of the compressed image can be set manually.

The first aggregated windowed self-attention learning layer consists of:

The pre-processed image processed by the patch partitioning transform layer is divided into 4×4 window, and then individually /> Self-attention learning is performed within each window; a separate /> After learning the windows, each window is shifted to the four corners by two adjacent pixels to form an aggregation window on the top corner of each window. Aggregation self-attention learning is then performed separately inside the aggregation window to strengthen the feature learning of the pixels of the back feature key points that may be included in the initial feature map, and output a first-level feature map.

(3.2) Secondary feature map extraction module: The primary feature map is input into the first patch merging layer and the second aggregated window self-attention learning layer, and the secondary feature map is output;

The first patch merging layer includes: performing non-convolutional downsampling on the first-level feature map obtained by the first-level feature map extraction module, and reducing the size of the first-level feature map to ;

The second aggregated window self-attention learning layer includes: dividing the first-level feature map into 4×4 window, and then individually /> Self-attention learning is performed within each window; a separate /> After learning the windows, each window is shifted to the four corners by two adjacent pixels to form an aggregation window on the top corner of each window. Aggregation self-attention learning is then performed separately inside the aggregation window to strengthen the feature learning of the pixels that may contain the back feature key points in the first-level feature map, and output the second-level feature map.

(3.3) Three-level feature map extraction module: The second-level feature map is input into the second patch merging layer and the third aggregated window self-attention learning layer, and the third-level feature map is output;

The second patch merging layer includes: performing non-convolutional downsampling on the secondary feature map obtained by the secondary feature map extraction module, and reducing the size of the secondary feature map to ;

The third aggregated window self-attention learning layer: divide the secondary feature map into 4×4 window, and then individually /> Self-attention learning is performed within each window; a separate /> After learning the windows, each window is shifted to the four corners by two adjacent pixels to form an aggregate window on the top corner of each window. Aggregate self-attention learning is then performed separately inside the aggregate window to strengthen the feature learning of the pixels that may contain the back feature key points in the secondary feature map, and output the third-level feature map.

(3.4) Four-level feature map extraction module: The three-level feature map is input into the third patch merging layer and the fourth aggregated window self-attention learning layer, and the four-level feature map is output;

The third patch merging layer includes: performing non-convolutional downsampling on the three-level feature map obtained by the three-level feature map extraction module, and converting the size of the three-level feature map into ;

The fourth aggregated window self-attention learning layer includes: dividing the three-level feature map into 4×4 window, and then individually /> Self-attention learning is performed within each window; a separate /> After learning the windows, each window is shifted to the four corners by two adjacent pixels to form an aggregation window on the top corner of each window. Aggregation self-attention learning is then performed separately inside the aggregation window to strengthen the feature learning of the pixels that may contain the back feature key points in the three-level feature map, and output a four-level feature map.

Non-convolutional downsampling includes:

When the feature map is input to the next-level feature extraction module, it will first enter the patch merging layer in the next-level module. The patch merging layer divides each 2×2 adjacent pixel area into a patch, and combines the pixels with the same value in each patch to obtain N 2×2 single-channel feature maps. Then, these 4 single-channel feature maps are spliced in the depth direction. Finally, a fully connected layer is used to perform a linear transformation in the depth direction of the single-channel feature map, thereby changing the depth of the single-channel feature map from C to 2C.

Self-attention learning includes: taking the vector corresponding to the information of each pixel in the single-channel feature map as the input vector, 、/> 、/> Generate the corresponding /> 、/> 、/> ,/> 、/> 、/> The vector length and the depth of the single channel feature map/> Keep consistent, corresponding to all pixels generated /> The process is as follows:

;

Generate corresponding to all pixels The process is as follows:

;

Generate corresponding to all pixels The process is as follows:

;

in, To make the length of/> , width is/> , depth is/> The matrix obtained by splicing all pixels in the single channel feature map together, /> 、/> 、/> To generate the corresponding /> 、/> 、/> The required weight matrix is 、/> 、/> The input vectors are respectively 、/> 、/> The information matrix obtained after transformation.

The income 、/> 、/> Information matrix input aggregation self-attention learning function:

;

Where D is the vector dimension, and B is the aggregate absolute offset between vectors;

.

Aggregate self-attention learning: Randomly take a 4×4 window and set the top left aggregation window Row, No./> The feature matrix of the column aggregation window pixels is/> , then the feature matrix of the pixels aggregated inside the window is/> With/> , the aggregated self-attention learning process is as follows:

;

in, For/> The back feature quantity in ,/> For/> The back feature value in the default /> The back feature value is 1.

Pass the above formula through The function will be normalized to obtain the characteristic matrix of the aggregation window: ; Repeat the above process to obtain the feature matrix of the remaining three aggregated window pixels in the same 4×4 window. When/> When , it is determined that the aggregation window only has back region features and no background region features, and the aggregation window is screened out; when it satisfies/> The back feature quantity in is given by/> Growth to/> , and it is the only maximum value among the feature quantities of the four aggregation windows, then it is determined that the aggregation window is the area in the four aggregation windows that is most likely to have the back feature key points; finally, the actual position of the back feature key points is obtained based on the obtained aggregation window feature information that is most likely to have the back feature key points.

In this embodiment, since the key feature points of the human back to be detected first are located at the junction of the human back image and the background image, even if the depth compensation map is obtained, it will be interfered by the background image pixels at the junction to a certain extent, resulting in deviations in the detection results of the back feature key points, and thus affecting the accuracy of the entire back acupuncture point positioning; therefore, four aggregated window self-attention learning layers are used to continuously iterate and strengthen the feature learning of the back feature key point pixels that may be contained in the feature maps at all levels.

As shown in Figure 5, this is the result of the detection of the key feature points of the human back without the aggregated window self-attention learning. It can be clearly seen that the key points in the upper left corner and the upper right corner of the back are biased towards the inner side of the back, which makes it impossible to accurately locate the positions of the acupoints close to the edge of the back, such as the Chengfeng acupoint; and the two key points at the bottom of the back are not detected due to the interference of the background image pixels at the junction; as shown in Figure 6, this is the result of the detection of the key feature points of the human back after the aggregated window self-attention learning. It can be clearly seen that the four key feature points of the back are close to the edge of the back, and the quadrilateral formed by the four points includes all 27 acupoints that need to be detected on the back, ensuring that the 27 acupoints can be correctly and accurately located through the back acupoint positioning formula.

(3) Input the four-level feature map into the back feature key point detection network to detect the back feature key points of the preprocessed back image, and obtain the specific coordinates of the back acupoints through the back acupoint positioning formula. The back feature key point detection network includes:

Global pooling layer: The obtained four-level feature map is globally averaged and pooled. The four-level feature map is compressed into a feature vector of fixed length. This feature vector is the feature representation of the back feature key points in the preprocessed image.

Classifier: The feature vector obtained after global average pooling is input into the fully connected layer to obtain the predicted labels and coordinates of the feature key points on the back of the image, and annotate them in the original back RGB image.

The back acupuncture point positioning formula includes:

default setting is the coordinate of the key point on the upper left of the back of the human body, /> is the coordinates of the key point on the upper right side of the human back, /> is the coordinate of the key point on the lower left side of the human back, /> is the coordinate of the key point on the lower right side of the back of the human body. The intersection of the optical axis of the deep binocular camera and the original back RGB image is the relative origin of the coordinate axis. According to the traditional Chinese medicine bone degree positioning method, the length between the two ends of the joint is measured as a certain inch, and the inch is converted into a mathematical distance. One inch is about 3.3 cm, which is used to determine the coordinates of the back acupuncture points. Since there are many acupuncture points on the back of the human body, and the acupuncture points are symmetrical about the back spine, 27 acupuncture points on the left half of the back and the midline of the back are selected for positioning:

(1) Riding the Wind: , Qu Heng:/> , Tianzong:/> ;

(2) Attached points: , Pohu:/> , paste blind:/> , Temple:/> , joke: , diaphragm:/> , Soul Gate: , Yang Gang:/> , intention: ; Air door: /> , Lung Shu: , Queyinshu:/> , Xinyu: , Du Yu:/> , diaphragm: , Wei Wan Xia Shu:/> , Liver Shu: , Gallbladder:/> , Spleen Shu: ;

(3) Body column: , Shinto:/> , Lingtai: , to Yang:/> , muscle contraction:/> ;

The following is a detailed description of the conversion of the above acupoints into coordinates:

(1) Chengfeng, Quheng and Tianzong are located in the shoulders;

(2) Fufen, Pohu, Gaomiang, Shentang, Xuxi, Geguan are located 3 inches below the spinous processes of the 2nd to 5th thoracic vertebrae; Hunmen, Yanggang, Yishe are located from the 2nd lumbar vertebrae to the 5th sacral vertebrae, 3 inches below each spinous process; Fengmen, Feishu, Queyinshu, Xinshu, Dushu, Geshu, Weiwanxiashu are located 1.5 inches below the spinous processes of the 2nd to 5th thoracic vertebrae; Ganshu, Danshu, Pishu are located from the 2nd lumbar vertebrae to the 5th sacral vertebrae, 1.5 inches below each spinous process;

(3) Body pillar, Shendao, Lingtai, Zhiyang, and Jinxu are located on the midline of the back.

The robot processor controls the robot's end effector to reach the acupuncture points on the back for kneading massage.

Claims (10)

1. An acupoint positioning method based on an aggregated window self-attention mechanism, characterized by comprising the following steps: (1) Acquire a human back image and perform preprocessing to obtain a preprocessed back image; (2) inputting the preprocessed back image into a multi-scale feature extraction network to obtain a back feature map; the multi-scale feature extraction network includes a plurality of feature map extraction modules, the preprocessed back image is sequentially processed by the plurality of feature map extraction modules, each feature map extraction module includes an aggregated window self-attention learning layer, and the aggregated window self-attention learning layer is used to strengthen feature learning of pixels containing feature key points in the extracted feature map; (3) Inputting the back feature map into the back feature key point detection network to detect the back feature key points and obtain the back feature key points; (4) Based on the key points of the back features, the specific coordinates of the back acupoints are obtained through the back acupoint positioning formula to achieve back acupoint positioning. 2. The acupoint locating method according to claim 1, characterized in that preprocessing the human back image comprises: Obtaining depth information of the human back image; The depth information is fused with the human back image through the depth information fusion formula to obtain a visual depth information map; The binary image of the human back is used as a mask to compensate for the lack of visual depth information map and obtain the preprocessed back image. 3. The acupoint locating method according to claim 2, characterized in that the step of obtaining the depth information of the human back image specifically comprises: obtaining a disparity map of the human back image, converting the disparity map into a depth map by a depth conversion formula to obtain the depth information, and the depth information calculation formula is: ; in, is the depth information, /> is parallax, /> is the baseline length, /> is the focal length, /> 、/> They are the horizontal coordinates of the same positions in the disparity map corresponding to the left and right cameras of the depth binocular camera. 4. The acupoint positioning method according to claim 3, characterized in that the depth information fusion formula is: ; in, is the depth information of the i-th pixel in the human back image, /> is the horizontal coordinate of the ith pixel in the human back image, /> is the ordinate of the i-th pixel in the human back image, /> is the horizontal coordinate of the intersection of the camera optical axis and the imaging plane, /> is the ordinate of the intersection of the camera optical axis and the imaging plane, /> 、/> 、/> is the fusion point coordinate of the i-th pixel in the visual depth information map, /> 、/> They are respectively/> Axis direction and /> The length of a single pixel in the axis direction. 5. The acupoint locating method according to claim 2, characterized in that the preprocessed back image is a human back image. Binarized image of and visual depth information map/> Use binary degree number/> The random fusion is obtained, and the fusion formula is: ; in, is the regression reward model,/> is an infinite non-repeating decimal, /> For the degree of reward. 6. The acupoint locating method according to claim 1 is characterized in that the multi-scale feature extraction network comprises a primary feature map extraction module, a secondary feature map extraction module, a tertiary feature map extraction module, and a quaternary feature map extraction module; the preprocessed back image is input into the primary feature map extraction module, and the primary feature map is output; the primary feature map extraction module comprises a patch partitioning transformation layer and a first aggregated window self-attention learning layer; the patch partitioning transformation layer comprises patch cropping and depth compression of the preprocessed back image; The primary feature map is input into a secondary feature map extraction module, and a secondary feature map is output, wherein the secondary feature map extraction module includes a first patch merging layer and a second aggregated window self-attention learning layer; The secondary feature map is input into a third-level feature map extraction module, and a third-level feature map is output. The third-level feature map extraction module includes a second patch merging layer and a third aggregated window self-attention learning layer; The three-level feature map is input into a four-level feature map extraction module, and a four-level feature map, i.e., a back feature map, is output. The four-level feature map extraction module includes a third patch merging layer and a fourth aggregated window self-attention learning layer; the first patch merging layer, the second patch merging layer, and the third patch merging layer are used to perform non-convolutional downsampling on the feature map. 7. The acupoint locating method according to claim 6, characterized in that the first aggregated window self-attention learning layer, the second aggregated window self-attention learning layer, the third aggregated window self-attention learning layer, and the fourth aggregated window self-attention learning layer all comprise the following steps: Divide the input feature map into windows to obtain several windows; Perform self-attention learning on each window separately; Shift each window after self-attention learning to the four corners to form an aggregate window on the corners of each window; Aggregate self-attention learning is performed inside the aggregation window, and the enhanced feature map is output. 8. The acupoint locating method according to claim 7, characterized in that the self-attention learning is to use the vector corresponding to the information of each pixel in the single-channel feature map as the input vector, generate the corresponding information matrix through the weight matrix, and input the information matrix into the self-attention learning function; The formula for the aggregated self-attention learning process is: ; in, The top left corner of the aggregation window/> Row, No./> The feature matrix of the column-aggregated window pixels; /> With/> is the feature matrix of pixels aggregated inside the window; /> is the feature matrix/> The back feature quantity in ,/> is the feature matrix/> The back feature quantity in ; When it appears When , the aggregation window only has back region features and no background region features, then the aggregation window is screened out; When the characteristic matrix The back feature quantity of is given by/> Growth to/> And when it is the only maximum value among the feature quantities of the four aggregation windows, it is determined that the aggregation window is the area in the four aggregation windows that is most likely to have back feature key points. 9. The acupoint locating method according to claim 8 is characterized in that the back feature key point detection network includes a global pooling layer and a classifier; the global pooling layer performs global average pooling on the back feature map, and the back feature map is compressed into a feature vector of a fixed length. The classifier inputs the feature vector into a fully connected layer to obtain the predicted labels and coordinates of the back feature key points. 10. A robot using the acupoint locating method according to claim 1, characterized in that it comprises an image acquisition module, a processing module, a control module, and an end effector; The image acquisition module is used to acquire the image of the back of the human body; The processing module is used to preprocess the acquired human back image to obtain a preprocessed back image; input the preprocessed back image into a multi-scale feature extraction network to obtain a back feature map; input the back feature map into a back feature key point detection network to perform back feature key point detection to obtain back feature key points; according to the back feature key points, the specific coordinates of the back acupoints are obtained by the back acupoint positioning formula to realize back acupoint positioning; The control module is used to control the end effector to perform kneading massage on the located acupuncture points on the back.