A method and system for intelligent identification of medical annotation symbols on body surface and precise positioning of physiotherapy robots

Abstract

The application discloses a body surface medical annotation symbol intelligent recognition and physiotherapy robot precise positioning method and system, and belongs to the technical field of physiotherapy robot control. The method comprises the following steps: drawing an annotation symbol on a patient body surface image and setting a semantic parameter; collecting image and depth information and pre-processing; detecting the symbol category and position through a YOLOv9 network; calculating the geometric center of each symbol; converting the pixel coordinates into three-dimensional physical coordinates in a robot base coordinate system through back projection and hand-eye calibration; analyzing the semantics to generate a physiotherapy sequence, and automatically executing the physiotherapy sequence by the robot after trajectory planning and collision pre-checking. The application realizes the full-process intelligentization from manual annotation to automatic execution of the robot, the positioning accuracy is better than 2 mm, the efficiency and safety are improved, and the application can be widely used in the fields of physical therapy, traditional Chinese medical massage and sports rehabilitation.

Description

Technical Field

[0001] This invention belongs to the field of physiotherapy robot control technology, specifically a method and system for intelligent recognition of medical marking symbols on the body surface and precise positioning of physiotherapy robots. Background Technology

[0002] In the fields of physical therapy, traditional Chinese massage, and sports rehabilitation, the precise location of treatment points directly affects efficacy and patient safety. In traditional clinical practice, after completing the assessment, the attending physician typically uses a skin safety marker to draw geometric symbols (such as circles, crosses, and arrows) on the patient's skin to mark the target areas, treatment methods, and contraindicated areas for subsequent physical therapy. This hand-drawn annotation method is intuitive and flexible and is widely used in clinical practice. However, when transmitting information to automated physical therapy robot systems, the following significant technical bottlenecks exist: (1) Insufficient accuracy of manual visual positioning. Operators rely on the naked eye to judge the center of the symbol. Due to differences in viewing angle, lighting and experience, the positioning error is usually 5 to 20 mm, which is difficult to meet the needs of high-precision operations such as acupoint massage and spinal correction, and may lead to deviation of therapeutic effect or safety risks.

[0003] (2) The efficiency of human-machine information transmission is extremely low. At present, the robot system mainly relies on manually dragging the end effector or manually setting it point by point through the APP interface to obtain the target point, which is cumbersome and time-consuming. For complex physiotherapy plans with more than 10 marked points, a single positioning often takes more than 15 minutes, which seriously restricts the clinical promotion of physiotherapy robots.

[0004] (3) The semantic annotation cannot be parsed in a structured way. Hand-drawn symbols not only indicate position, but also carry operational semantics (such as force application area, avoidance area, direction indication, etc.). However, the existing system lacks the technical means to automatically identify the symbol category and map it to robot motion parameters, resulting in a break in information transmission.

[0005] (4) Lack of universal solutions across body surface areas. Existing computer-aided positioning solutions are mostly designed for specific parts (such as the back or neck), which are difficult to adapt to the morphological differences of multiple parts such as limbs, head and face, and have poor universality.

[0006] In recent years, deep learning object detection technology has developed rapidly. YOLOv9, as the latest representative architecture, introduces a programmable gradient information (PGI) mechanism and a generalized efficient layer aggregation network (Gelan-C), demonstrating excellent performance in small object detection in complex backgrounds (inference speed >60 FPS, leading mAP), and has the potential for real-time physiotherapy navigation. However, there is currently no complete technical solution for applying YOLOv9 to multi-class recognition of medical annotation symbols on the body surface and further integrating it into the one-click localization and automatic execution system of physiotherapy robots. Summary of the Invention

[0007] To solve the above problems, the technical solution provided by the present invention is as follows: The present invention discloses a method for intelligent recognition of medical annotation symbols on the body surface and precise positioning of a physiotherapy robot, the method comprising the following steps. S1. Consultation and annotation stage: Consultants use mobile devices to draw medical annotation symbols on the patient's body surface images and set operation semantics and treatment parameters for each symbol; S2, Image Acquisition and Preprocessing Stage: Acquire color images and depth information of the patient's body surface, and perform contrast enhancement and depth filtering preprocessing. S3, Multi-class symbol detection stage: Real-time target detection is performed on the preprocessed image using a neural network model trained on a specialized dataset, and the symbol category label, bounding box coordinates, and confidence score are output. S4, Geometric Center Point Precise Calculation Stage: Calculate the geometric center pixel coordinates (u*, v*) for each symbol; S5, 3D coordinate mapping stage: Using the camera intrinsic parameter matrix K and the depth value d, the center point pixel coordinates are transformed to the camera coordinate system through the pinhole camera back projection model, and then transformed to the robot base coordinate system through the hand-eye calibration extrinsic parameter matrix T_cam_to_base to obtain millimeter-level 3D physical coordinates; S6. Motion Planning and Execution Phase: The robot generates an ordered sequence of physiotherapy points by parsing and annotating the semantics, performs inverse kinematics solving, fifth-order polynomial trajectory planning, and full-path collision pre-detection. After confirmation by the operator's AR visualization, the robot automatically executes all physiotherapy operations, with real-time collision monitoring and force control feedback throughout the process.

[0008] Preferably, the neural network model trained on the specific dataset in step S2 is specifically: A specialized dataset of medical annotation symbols for the body surface was constructed and a neural network model was developed to obtain the neural network model. Includes the following steps (1) Raw data collection; (2) Data labeling; (3) Pre-trained model; (4) Model optimization; (5) Verification and evaluation.

[0009] Preferably, step (3) specifically involves: Pre-trained YOLOv9-C weights on the dataset are used as the starting point for transfer learning, and the transfer learning strategy is as follows: Freeze strategy: During training rounds 1 to 50 (warm-up phase), freeze the weights of the first 10 layers of the backbone network Gelan-C, and only fine-tune the neck FPN+PAN layer and the detection head, with a learning rate lr0=0.001; Full network fine-tuning: In rounds 51-300, all layers are unfrozen, the entire network participates in gradient updates, and the learning rate decays from 0.01 to 0.0001 using a cosine annealing strategy; Data augmentation configuration: Enable Mosaic enhancement (4 images stitched together, probability = 1.0), random horizontal / vertical flip (probability = 0.5), random rotation (±15°), HSV color gamut perturbation (h = 0.015, s = 0.7, v = 0.4), random noise and blur; Training hyperparameters: batch_size=16, input size 640×640, optimizer SGD (momentum=0.937, weight_decay=0.0005), training on an NVIDIA A100 GPU took approximately 4 hours; The total training loss is calculated using the PGI auxiliary branch weighted joint loss: L_total = L_main + λ × L_aux; In the formula, L_main is the multi-task loss of the main detection head, L_aux is the supervision loss of the PGI auxiliary invertible branch, and λ is the balancing weight (default value is 0.25). Both losses are backpropagated to the shared backbone feature extraction network. The main head loss L_main consists of three parts: L_main = L_cls + L_box + L_dfl L_cls = Σ [-y·log(p) - (1-y)·log(1-p)] (Binary cross-entropy classification loss) L_box = 1 - CIoU(b_pred, b_gt) (Full IoU bounding box regression loss) L_dfl = Σ [-(y_i+1-y)·log(S_i) - (y-y_i)·log(S_{i+1})] (Distribution focus loss) In the formula, L_cls is the binary cross-entropy classification loss, y is the ground truth class label, and p is the predicted probability; L_box uses CIoU (complete IoU) loss, which comprehensively considers the overlap rate, center point distance, and aspect ratio; L_dfl is the distribution focus loss, which models the bounding box coordinates as a discrete probability distribution to improve the fine-grained positioning accuracy; b_pred and b_gt are the predicted box and the ground truth box, respectively; S_i is the Softmax probability output.

[0010] Preferably, step S5 specifically includes S51, Camera intrinsic parameter matrix K (pinhole camera model): In the formula, fx and fy are the focal lengths of the camera in the x / y directions (in pixels), and cx and cy are the coordinates of the optical principal point (in pixels), which are precisely determined by the checkerboard calibration program, with a reprojection error of no more than 0.5 pixels; S52, Pixel coordinates to camera 3D coordinates back projection: Xc = (u* - cx) × d / fx Yc = (v* - cy) × d / fy Zc = d In the formula, (u*,v*) are the center point pixel coordinates output in step S3; d is the depth measurement value of the depth sensor at this pixel position (in mm, which needs to be registered with the RGB image for depth-color); (Xc,Yc,Zc) are the three-dimensional coordinates of the center point in the camera coordinate system, in mm; S53. Transformation from camera coordinate system to robot base coordinate system: P_base = T_cam_to_base × [Xc, Yc, Zc, 1]ᵀ In the formula, T_cam_to_base is the 4×4 homogeneous transformation matrix obtained by hand-eye calibration in step S1.2, which includes a 3×3 rotation submatrix R and a 3×1 translation vector t; P_base = [Xw,Yw,Zw,1]ᵀ is the three-dimensional physical coordinate of the center point in the robot base coordinate system, in mm, with an accuracy better than 2mm; S54. Comprehensive Error Analysis: The final three-dimensional positioning error of the system is synthesized from the root mean square of various errors: σ_total = √(σ_detect² + σ_center² + σ_depth² + σ_calib²) In the formula, σ_detect≈1px (YOLOv9 detection box error), σ_center≈0.5px (geometric center algorithm error), σ_depth≈1mm (depth sensor error), and σ_calib≈0.3mm (calibration residual); the synthesized accuracy σ_total <2mm, which meets the accuracy requirements of clinical physiotherapy.

[0011] Preferably, step S6 specifically includes S61, Annotation Semantic Parsing: The system reads the structured annotation task file output in step S1, parses the category semantics of each point and the corresponding operation parameters, and generates an ordered sequence of physiotherapy points; S62. Inverse Kinematics Solution: For each target point P_base in the ordered sequence of therapeutic points, combined with the end-effector approach direction vector n (provided by the symbol category or arrow direction), calculate the robot joint angle target configuration q_tgt: q_tgt = IK(P_base, n, q_seed) In the formula, IK(·) is the robot inverse kinematics solution function (using numerical iteration or analytical method, depending on the robot model); P_base is the three-dimensional coordinates of the target point; n is the end-effector approach direction vector; q_seed is the initial seed joint angle (taking the current joint angle), used to select the optimal configuration in the case of multiple solutions (minimum joint motion criterion). S63. Trajectory Planning and Collision Pre-detection: Before the robot's actual movement, fifth-order polynomial interpolation is used to plan the motion trajectory from the current configuration q_curr to the target configuration q_tgt in joint space, and full-path collision pre-detection is performed. q(t)=a3+a1t+a2t5+a3t5+a4t5+a4t5; In the formula, the coefficients a0, a1, a2, a3, and a4 are uniquely determined by the six boundary conditions of the starting and ending points, velocity (=0), and acceleration (=0), ensuring smooth start and stop of the robot; S64. Visualization and Operator Confirmation: After the trajectory planning is completed, the system displays it in an augmented reality overlay on the operation interface and waits for operator confirmation. S65. Automatic execution and real-time monitoring: The robot automatically visits each physiotherapy point in sequence according to the planned path and performs the corresponding operation actions.

[0012] A system for intelligent recognition of medical marking symbols on the body surface and precise positioning of a physiotherapy robot, the system being used to execute the above-described method, the system comprising: The mobile annotation module allows medical personnel to draw medical annotation symbols on images of patients' bodies, set operational semantics and treatment parameters for each symbol, and generate structured annotation task files. The image acquisition module is used to acquire color images and depth information of the patient's body surface, and to perform contrast enhancement and depth filtering preprocessing on the acquired images; The multi-class symbol detection module uses a neural network model trained on a specialized dataset to perform real-time target detection on preprocessed images and output symbol category labels, bounding box coordinates, and confidence scores. The geometric center accurate calculation module is used to execute category-specific geometric algorithms based on the detected symbol category and bounding box to accurately calculate the pixel coordinates of the geometric center of each symbol. The 3D coordinate mapping module is used to transform the center point pixel coordinates to the camera coordinate system through the pinhole camera back projection model using the camera intrinsic parameter matrix and depth value, and then transform them to the robot base coordinate system through the hand-eye calibration extrinsic parameter matrix to obtain the 3D physical coordinates. The semantic parsing and motion planning module is used to parse the annotation semantics in the structured annotation task file, generate an ordered sequence of physiotherapy points, and perform inverse kinematics solving, trajectory planning, and full-path collision pre-detection. The augmented reality visualization confirmation module is used to overlay and display the detected symbol points, the planned motion path of the robot end effector, and the operating parameters of each physiotherapy point on the operation interface in an augmented reality manner, so that the operator can confirm or adjust the execution plan. The robot execution and safety monitoring module is used to control the robot to automatically perform physiotherapy operations according to the planned path, and to continuously perform real-time collision monitoring and force control feedback during the execution process.

[0013] Preferably, the multi-class symbol detection module uses a Gelan-C generalized efficient layer aggregation network as the backbone feature extraction network, and is trained by combining a programmable gradient information-assisted reversible branch mechanism; the programmable gradient information-assisted branch only participates in gradient calculation during the training phase and does not participate in calculation during the inference phase.

[0014] Preferably, the geometric center accurate calculation module includes: The circle symbol calculation submodule is used to calculate the geometric center of the circle symbol using the least squares circle fitting method; The cross symbol calculation submodule is used to calculate the geometric center of the cross symbol using the double-stroke PCA principal axis intersection method; The triangle symbol calculation submodule is used to calculate the geometric center of a triangle symbol using the centroid method of three vertices; The star symbol calculation submodule is used to calculate the geometric center of the star symbol using the convex hull center method; The square symbol calculation submodule is used to calculate the geometric center of the square symbol using the minimum bounding rectangle center method; The directional arrow symbol calculation submodule is used to calculate the coordinates of the force application point and the direction angle of the directional arrow symbol using the skeleton root positioning and direction angle extraction method.

[0015] Preferably, the semantic parsing and motion planning module inputs the three-dimensional coordinates of the cross-type avoidance symbol as an obstacle constraint into the collision pre-detection module, and automatically sets a safe distance constraint for each link of the robot for the avoidance point during the trajectory planning stage, so as to ensure that the robot maintains a safe gap of not less than 30mm with the avoidance area during the execution process. The square symbol calculation submodule is also used to automatically generate a raster scanning path based on the major axis direction, minor axis direction and size of the smallest bounding rectangle. The robot end effector performs parallel reciprocating scanning in rows within the square area, and the row spacing is automatically calculated based on the diameter of the end tool.

[0016] Preferably, the mobile annotation module supports gesture drawing mode and symbol toolbar selection mode, and the drawn annotation information is transmitted to the semantic parsing and motion planning module in a structured JSON file format; the robot execution and safety monitoring module performs rolling verification of the collision risk of the future predicted position at a frequency of not less than 10Hz and updates the execution progress in real time; the operator can stop the robot movement at any time through the emergency stop button. Beneficial effects

[0017] Compared with the prior art, the technical solution provided by this invention has the following advantages: This invention discloses a method and system for intelligent recognition of medical annotation symbols on the body surface and precise positioning of a physiotherapy robot. The method includes: a patient drawing medical annotation symbols on a patient's body surface image using a mobile device and setting operational semantics and treatment parameters; acquiring color images and depth information of the patient's body surface and performing preprocessing; performing real-time target detection using a neural network model trained on a specialized dataset, outputting symbol category labels, bounding box coordinates, and confidence scores; calculating the geometric center pixel coordinates of each symbol; transforming the center point pixel coordinates to the camera coordinate system using a pinhole camera back-projection model with camera intrinsic parameters and depth values, and then transforming them to the robot's base coordinate system using a hand-eye calibration extrinsic parameter matrix to obtain three-dimensional physical coordinates; parsing the annotation semantics to generate an ordered sequence of physiotherapy points, performing inverse kinematics solving, trajectory planning, and full-path collision pre-detection; and automatically performing physiotherapy operations after AR visualization confirmation. This invention integrates deep learning target detection, category-specific geometric center calculation, hand-eye calibration 3D mapping, and robot motion planning and control, realizing the intelligentization of the entire process from manual drawing to automatic execution by the robot for medical annotation symbols on the body surface. It significantly improves the positioning accuracy, operational efficiency, and clinical safety of physiotherapy robots and can be widely applied in the fields of physical therapy, traditional Chinese massage, and sports rehabilitation. Attached Figure Description

[0018] Figure 1 This is a mapping diagram of the six types of medical annotation symbols on the body surface and their recognition details with robot motion, as presented in this invention. Figure 2 This is a flowchart of a method for intelligent recognition of medical annotation symbols on the body surface and precise positioning of a physiotherapy robot according to the present invention. Detailed Implementation

[0019] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort should fall within the scope of protection of the present application.

[0020] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate for the embodiments of this application described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0021] In this application, the terms "upper," "lower," "left," "right," "front," "rear," "top," "bottom," "inner," "outer," "middle," "vertical," "horizontal," "lateral," and "longitudinal" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. These terms are primarily for the purpose of better describing this application and its embodiments, and are not intended to limit the indicated device, element, or component to having a specific orientation, or to be constructed and operated in a specific orientation.

[0022] Furthermore, in addition to indicating location or positional relationship, some of the aforementioned terms may also have other meanings. For example, the term "above" may also be used in some cases to indicate a certain dependency or connection relationship. Those skilled in the art can understand the specific meaning of these terms in this application based on the specific circumstances.

[0023] Furthermore, the terms "installation," "setup," "equipped with," "connection," "linking," and "socketing" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral structure; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium, or an internal connection between two devices, components, or parts. Those skilled in the art can understand the specific meaning of these terms in this application based on the specific circumstances.

[0024] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments. Example 1

[0025] See attached document Figure 1-2 This embodiment discloses a method for intelligent recognition of medical annotation symbols on the body surface and precise positioning of a physiotherapy robot, the method comprising the following steps. S1. Consultation and annotation stage: Consultants use mobile devices to draw medical annotation symbols on the patient's body surface images and set operation semantics and treatment parameters for each symbol; S2, Image Acquisition and Preprocessing Stage: Acquire color images and depth information of the patient's body surface, and perform contrast enhancement and depth filtering preprocessing. S3, Multi-class symbol detection stage: Real-time target detection is performed on the preprocessed image using a neural network model trained on a specialized dataset, and the symbol category label, bounding box coordinates, and confidence score are output. S4, Geometric Center Point Precise Calculation Stage: Calculate the geometric center pixel coordinates (u*, v*) for each symbol; S5, 3D coordinate mapping stage: Using the camera intrinsic parameter matrix K and the depth value d, the center point pixel coordinates are transformed to the camera coordinate system through the pinhole camera back projection model, and then transformed to the robot base coordinate system through the hand-eye calibration extrinsic parameter matrix T_cam_to_base to obtain millimeter-level 3D physical coordinates; S6. Motion Planning and Execution Phase: The robot generates an ordered sequence of physiotherapy points by parsing and annotating the semantics, performs inverse kinematics solving, fifth-order polynomial trajectory planning, and full-path collision pre-detection. After confirmation by the operator's AR visualization, the robot automatically executes all physiotherapy operations, with real-time collision monitoring and force control feedback throughout the process.

[0026] The neural network model trained on the specific dataset in step S2 is specifically as follows: A specialized dataset of medical annotation symbols for the body surface was constructed and a neural network model was developed to obtain the neural network model. Includes the following steps (1) Raw data collection; (2) Data labeling; (3) Pre-trained model; (4) Model optimization; (5) Verification and evaluation.

[0027] Preferably, step (3) specifically involves: Pre-trained YOLOv9-C weights on the dataset are used as the starting point for transfer learning, and the transfer learning strategy is as follows: Freeze strategy: During training rounds 1 to 50 (warm-up phase), freeze the weights of the first 10 layers of the backbone network Gelan-C, and only fine-tune the neck FPN+PAN layer and the detection head, with a learning rate lr0=0.001; Full network fine-tuning: In rounds 51-300, all layers are unfrozen, the entire network participates in gradient updates, and the learning rate decays from 0.01 to 0.0001 using a cosine annealing strategy; Data augmentation configuration: Enable Mosaic enhancement (4 images stitched together, probability = 1.0), random horizontal / vertical flip (probability = 0.5), random rotation (±15°), HSV color gamut perturbation (h = 0.015, s = 0.7, v = 0.4), random noise and blur; Training hyperparameters: batch_size=16, input size 640×640, optimizer SGD (momentum=0.937, weight_decay=0.0005), training on an NVIDIA A100 GPU took approximately 4 hours; The total training loss is calculated using the PGI auxiliary branch weighted joint loss: L_total = L_main + λ × L_aux; In the formula, L_main is the multi-task loss of the main detection head, L_aux is the supervision loss of the PGI auxiliary invertible branch, and λ is the balancing weight (default value is 0.25). Both losses are backpropagated to the shared backbone feature extraction network. The main head loss L_main consists of three parts: L_main = L_cls + L_box + L_dfl L_cls = Σ [-y·log(p) - (1-y)·log(1-p)] (Binary cross-entropy classification loss) L_box = 1 - CIoU(b_pred, b_gt) (Full IoU bounding box regression loss) L_dfl = Σ [-(y_i+1-y)·log(S_i) - (y-y_i)·log(S_{i+1})] (Distribution focus loss) In the formula, L_cls is the binary cross-entropy classification loss, y is the ground truth class label, and p is the predicted probability; L_box uses CIoU (complete IoU) loss, which comprehensively considers the overlap rate, center point distance, and aspect ratio; L_dfl is the distribution focus loss, which models the bounding box coordinates as a discrete probability distribution to improve the fine-grained positioning accuracy; b_pred and b_gt are the predicted box and the ground truth box, respectively; S_i is the Softmax probability output.

[0028] Preferably, step S5 specifically includes S51, Camera intrinsic parameter matrix K (pinhole camera model): In the formula, fx and fy are the focal lengths of the camera in the x / y directions (in pixels), and cx and cy are the coordinates of the optical principal point (in pixels), which are precisely determined by the checkerboard calibration program, with a reprojection error of no more than 0.5 pixels; S52, Pixel coordinates to camera 3D coordinates back projection: Xc = (u* - cx) × d / fx Yc = (v* - cy) × d / fy Zc = d In the formula, (u*,v*) are the center point pixel coordinates output in step S3; d is the depth measurement value of the depth sensor at this pixel position (in mm, which needs to be registered with the RGB image for depth-color); (Xc,Yc,Zc) are the three-dimensional coordinates of the center point in the camera coordinate system, in mm; S53. Transformation from camera coordinate system to robot base coordinate system: P_base = T_cam_to_base × [Xc, Yc, Zc, 1]ᵀ In the formula, T_cam_to_base is the 4×4 homogeneous transformation matrix obtained by hand-eye calibration in step S1.2, which includes a 3×3 rotation submatrix R and a 3×1 translation vector t; P_base = [Xw,Yw,Zw,1]ᵀ is the three-dimensional physical coordinate of the center point in the robot base coordinate system, in mm, with an accuracy better than 2mm; S54. Comprehensive Error Analysis: The final three-dimensional positioning error of the system is synthesized from the root mean square of various errors: σ_total = √(σ_detect² + σ_center² + σ_depth² + σ_calib²) In the formula, σ_detect≈1px (YOLOv9 detection box error), σ_center≈0.5px (geometric center algorithm error), σ_depth≈1mm (depth sensor error), and σ_calib≈0.3mm (calibration residual); the synthesized accuracy σ_total <2mm, which meets the accuracy requirements of clinical physiotherapy.

[0029] Step S6 specifically includes S61, Annotation Semantic Parsing: The system reads the structured annotation task file output in step S1, parses the category semantics of each point and the corresponding operation parameters, and generates an ordered sequence of physiotherapy points; S62. Inverse Kinematics Solution: For each target point P_base in the ordered sequence of therapeutic points, combined with the end-effector approach direction vector n (provided by the symbol category or arrow direction), calculate the robot joint angle target configuration q_tgt: q_tgt = IK(P_base, n, q_seed) In the formula, IK(·) is the robot inverse kinematics solution function (using numerical iteration or analytical method, depending on the robot model); P_base is the three-dimensional coordinates of the target point; n is the end-effector approach direction vector; q_seed is the initial seed joint angle (taking the current joint angle), used to select the optimal configuration in the case of multiple solutions (minimum joint motion criterion). S63. Trajectory Planning and Collision Pre-detection: Before the robot's actual movement, fifth-order polynomial interpolation is used to plan the motion trajectory from the current configuration q_curr to the target configuration q_tgt in joint space, and full-path collision pre-detection is performed. q(t)=a3+a1t+a2t5+a3t5+a4t5+a4t5; In the formula, the coefficients a0, a1, a2, a3, and a4 are uniquely determined by the six boundary conditions of the starting and ending points, velocity (=0), and acceleration (=0), ensuring smooth start and stop of the robot; S64. Visualization and Operator Confirmation: After the trajectory planning is completed, the system displays it in an augmented reality overlay on the operation interface and waits for operator confirmation. S65. Automatic execution and real-time monitoring: The robot automatically visits each physiotherapy point in sequence according to the planned path and performs the corresponding operation actions. Example 2

[0030] This embodiment provides a system for intelligent recognition of medical annotation symbols on the body surface and precise positioning of a physiotherapy robot. The system is used to execute the method of Embodiment 1. The system includes... The mobile annotation module allows medical personnel to draw medical annotation symbols on images of patients' bodies, set operational semantics and treatment parameters for each symbol, and generate structured annotation task files. The image acquisition module is used to acquire color images and depth information of the patient's body surface, and to perform contrast enhancement and depth filtering preprocessing on the acquired images; The multi-class symbol detection module uses a neural network model trained on a specialized dataset to perform real-time target detection on preprocessed images and output symbol category labels, bounding box coordinates, and confidence scores. In an optional embodiment of the present invention, the detection backbone network adopts the YOLOv9-C structure, which includes the GELAN feature extraction network and the PGI-assisted training mechanism. Alternatively, YOLO series networks with similar architectures such as YOLOv8 and YOLOv7 can be used as alternatives, all of which are equivalent implementations of the present invention.

[0031] The geometric center accurate calculation module is used to execute category-specific geometric algorithms based on the detected symbol category and bounding box to accurately calculate the pixel coordinates of the geometric center of each symbol. The 3D coordinate mapping module is used to transform the center point pixel coordinates to the camera coordinate system through the pinhole camera back projection model using the camera intrinsic parameter matrix and depth value, and then transform them to the robot base coordinate system through the hand-eye calibration extrinsic parameter matrix to obtain the 3D physical coordinates. The semantic parsing and motion planning module is used to parse the annotation semantics in the structured annotation task file, generate an ordered sequence of physiotherapy points, and perform inverse kinematics solving, trajectory planning, and full-path collision pre-detection. The augmented reality visualization confirmation module is used to overlay and display the detected symbol points, the planned motion path of the robot end effector, and the operating parameters of each physiotherapy point on the operation interface in an augmented reality manner, so that the operator can confirm or adjust the execution plan. The robot execution and safety monitoring module is used to control the robot to automatically perform physiotherapy operations according to the planned path, and to continuously perform real-time collision monitoring and force control feedback during the execution process.

[0032] The multi-class symbol detection module uses a Gelan-C generalized efficient layer aggregation network as the backbone feature extraction network, and is trained by combining a programmable gradient information-assisted reversible branch mechanism; the programmable gradient information-assisted branch only participates in gradient calculation during the training phase and does not participate in calculation during the inference phase.

[0033] The precise geometric center calculation module includes: The circle symbol calculation submodule is used to calculate the geometric center of the circle symbol using the least squares circle fitting method; The cross symbol calculation submodule is used to calculate the geometric center of the cross symbol using the double-stroke PCA principal axis intersection method; The triangle symbol calculation submodule is used to calculate the geometric center of a triangle symbol using the centroid method of three vertices; The star symbol calculation submodule is used to calculate the geometric center of the star symbol using the convex hull center method; The square symbol calculation submodule is used to calculate the geometric center of the square symbol using the minimum bounding rectangle center method; The directional arrow symbol calculation submodule is used to calculate the coordinates of the force application point and the direction angle of the directional arrow symbol using the skeleton root positioning and direction angle extraction method.

[0034] Preferably, the semantic parsing and motion planning module inputs the three-dimensional coordinates of the cross-type avoidance symbol as an obstacle constraint into the collision pre-detection module, and automatically sets a safe distance constraint for each link of the robot for the avoidance point during the trajectory planning stage, so as to ensure that the robot maintains a safe gap of not less than 30mm with the avoidance area during the execution process. The square symbol calculation submodule is also used to automatically generate a raster scanning path based on the major axis direction, minor axis direction and size of the smallest bounding rectangle. The robot end effector performs parallel reciprocating scanning in rows within the square area, and the row spacing is automatically calculated based on the diameter of the end tool.

[0035] The mobile annotation module supports gesture drawing mode and symbol toolbar selection mode. The drawn annotation information is transmitted to the semantic parsing and motion planning module in a structured JSON file format. The robot execution and safety monitoring module performs rolling verification of the collision risk of the future predicted position at a frequency of not less than 10Hz and updates the execution progress in real time. The operator can stop the robot's movement at any time through the emergency stop button.

[0036] Compared with the prior art, the present invention has the following beneficial effects: (i) Significantly improve the positioning accuracy of surface marking symbols This invention achieves a 3D localization accuracy better than 2mm through a three-level accuracy guarantee mechanism: deep learning object detection, category-specific geometric center algorithm, and hand-eye calibration 3D mapping. First, a network model is used to detect medical annotation symbols on the body surface. Its introduced programmable gradient information (PGI) mechanism effectively solves the information bottleneck problem of losing key feature information during the forward propagation of deep neural networks, ensuring reliable feature extraction by preserving accurate gradients used to update weights. The Generalized Efficient Layer Aggregation Network (GELAN) maximizes parameter efficiency, enabling the model to achieve high-precision detection with fewer floating-point operations. Second, dedicated geometric center algorithms are designed for different symbol categories (circles use least-squares circle fitting, crosses use PCA principal axis intersections, and arrows use skeleton root localization, etc.), controlling the center localization error to within 0.5 pixels, improving accuracy by more than 4 times compared to directly taking the bounding box center (error 23 pixels). Finally, through rigorous checkerboard calibration (reprojection error ≤ 0.5 pixels) and hand-eye calibration, combined with the comprehensive error synthesis formula σ_total=√(σ_detect²+σ_center²+σ_depth²+σ_calib²), each error component is quantitatively controlled, achieving a synthesis accuracy better than 2mm. In contrast, traditional manual visual positioning errors typically reach 520mm. This invention improves positioning accuracy by more than an order of magnitude, fully meeting the positioning requirements of high-precision clinical physiotherapy operations such as acupoint massage and spinal correction.

[0037] (ii) Significantly improve the efficiency of information transmission from consultation to implementation of physiotherapy plans. This invention constructs a complete automated closed-loop chain from "diagnosis annotation—image acquisition—symbol detection—coordinate mapping—motion planning—automatic execution," completely changing the inefficient mode of traditional physiotherapy robots that rely on manual dragging of end effectors or manually setting target points point by point through an APP interface. The patient directly draws medical annotation symbols on the patient's body surface image using a mobile device and simultaneously sets the operational semantics and treatment parameters. The drawing information is transmitted to the robot control system in a structured JSON file format. Subsequently, the system automatically completes symbol detection, center calculation, 3D mapping, and motion planning. The operator only needs to make final confirmation through an AR visualization interface to start automatic execution. The entire process, from the end of the consultation to the start of robot execution, can be shortened from more than 15 minutes of traditional manual point-by-point setting to within a few minutes, improving efficiency by more than 10 times. For complex physiotherapy programs containing more than 10 annotation points, the efficiency advantage of this invention is particularly prominent, effectively solving the bottleneck problem of "extremely low efficiency of human-machine information transmission" in the background technology, and clearing the efficiency barrier for the large-scale clinical promotion of physiotherapy robots.

[0038] (III) Achieving structured parsing and automatic mapping of medical annotation symbol semantics This invention elevates traditional hand-drawn symbols from simple "visual positional marks" to "structured digital labels" carrying complete operational semantics. Specifically, the system automatically maps operational semantics based on symbol category: circle symbols are mapped to fixed-point pressing operations; cross symbols are automatically identified as obstacle avoidance (the system sets a safety distance constraint of no less than 30mm for them during the trajectory planning stage); square symbols automatically generate raster scanning paths based on the major and minor axes and dimensions of the smallest circumscribed rectangle (line spacing is automatically calculated based on the diameter of the end tool); and directional arrow symbols extract the coordinates of the arrow shaft root and the directional angle to determine the point and direction of force application. This "symbol as instruction" design allows clinicians to complete the complete programming of robot treatment plans without learning complex robot programming languages, simply by drawing symbols familiar to clinical practice. Compared to the existing technology where hand-drawn symbols only serve as visual marks and their operational semantics require secondary interpretation and conversion by humans, this invention fundamentally eliminates information transmission gaps and achieves seamless conversion between clinical medical language and robot control language.

[0039] (iv) Exhibits excellent robustness and real-time performance in complex surface environments for symbol detection. This invention addresses the unique characteristics of medical annotation symbols on the body surface—small symbol size (often <20×20 pixels), uneven surface lighting, interference from skin texture and hair, and blurred human body contours—through specific optimizations. At the data level, a dedicated dataset of medical annotation symbols on the body surface was constructed, and diverse data augmentation strategies were employed, including Mosaic enhancement (4-image stitching), random horizontal / vertical flipping, random rotation (±15°), HSV color gamut perturbation, random noise, and blurring, effectively improving the model's adaptability to complex clinical environments. At the model level, a transfer learning strategy was adopted (freezing the weights of the first 10 layers of the backbone network for the first 50 rounds to utilize pre-trained general features, and fine-tuning the entire network for the next 250 rounds to adapt to the special shapes of the symbols on the body surface), combined with a PGI-assisted reversible branch to provide rich gradient signals during the training phase to mitigate the loss of deep information. In terms of loss function design, the CIoU loss comprehensively considers overlap rate, center point distance, and aspect ratio, resulting in more stable convergence than traditional IoU; the distributed focus loss models the bounding box coordinates as a discrete probability distribution, improving boundary precision. Experimental results show that the signal-to-noise ratio of the symbol region is improved by about 6dB after preprocessing, the recall rate of small targets (such as a circle with a diameter of 5mm) is increased from 82% to 96%, the overall mAP@0.5 reaches 0.947, and the inference speed is 62 FPS, which ensures both real-time performance and detection accuracy.

[0040] (v) Ensure safety throughout the entire process of robotic physiotherapy operation. This invention constructs a multi-layered safety assurance system during the motion planning and execution phases. First, a full-path collision pre-detection is performed before trajectory planning. Each link of the robot is simplified into a cylindrical / spherical bounding box, and spatial intersection calculations are performed with known obstacles (including avoidance zones marked with crosses, treatment beds, etc.) to pre-mark unsafe paths. Second, a fifth-order polynomial is used to interpolate the trajectory in joint space. The coefficients are uniquely determined by six boundary conditions: start and end positions, velocity (=0), and acceleration (=0), ensuring continuous and impact-free acceleration for patient comfort. Third, during execution, the collision risk at the predicted future position is verified at a frequency of at least 10Hz, and the end-effector contact force is monitored in real time, with output torque adjusted through force control feedback. Fourth, the operator can stop the robot's movement at any time via an emergency stop button. These multiple safety mechanisms work together to reduce the actual collision risk to zero, avoiding medical safety accidents caused by improper paths or uncontrolled force.

[0041] (vi) Enhance operators' ability to conduct visual reviews and make decisions. This invention introduces an augmented reality (AR) visualization confirmation module into the user interface, overlaying detected symbol points, the planned motion path of the robot's end effector, and the operational parameters of each treatment point onto the patient's surface image in AR. Before execution, the operator can visually review the relative relationship between the robot's planned actions and the patient's surface, and fine-tune the target point position or modify parameters through gestures or mouse dragging. This design organically combines automated machine planning with human clinical judgment—leveraging the efficiency advantages of automation while retaining the indispensable human supervision in medical decision-making. Clinical testing shows that the average operator AR confirmation time is only 15 seconds, and the error rate (such as failure to detect path conflicts before startup) has decreased from 8% without AR confirmation to 0%. The AR visualization confirmation mechanism effectively compensates for the trust shortcomings of the "fully automated black box" mode in medical scenarios, enabling operators to complete the final review of the execution plan in a "what you see is what you get" manner, significantly improving the system's safety and clinical acceptability.

[0042] (vii) Possesses versatility across body surface regions and good scalability This invention does not customize designs for specific body parts (such as the back or neck), but rather establishes a universal technical framework: the mobile annotation module supports both gesture drawing and symbol toolbar selection modes, adaptable to annotation needs of any body part; the multi-category symbol detection module covers six common medical annotation symbols: circles, crosses, triangles, stars, squares, and directional arrows, and the model architecture supports expanding to new symbol categories through incremental training; the 3D coordinate mapping module is based on a universal pinhole camera model and hand-eye calibration matrix, independent of geometric priors for specific body parts; the semantic parsing and motion planning module decouples symbol semantics from robot motion parameters, allowing adaptation to different robot models through configuration files. Therefore, this invention can be widely applied to various clinical scenarios such as physical therapy, traditional Chinese massage, sports rehabilitation, and postoperative physiotherapy, and is compatible with different brands and configurations of physiotherapy robot platforms, exhibiting good versatility and scalability.

[0043] In summary, this invention organically integrates multiple technologies, including deep learning target detection, category-specific geometric center calculation, hand-eye calibration 3D mapping, AR visualization confirmation, and robot motion planning and control. It systematically solves the core problems of insufficient positioning accuracy, low information transmission efficiency, lack of semantic parsing, and weak safety assurance in the complete technical chain of medical annotation symbols on the body surface from manual drawing to automatic execution by robots. This provides a complete technical solution for the intelligentization, precision, and clinical application of physiotherapy robots.

[0044] The above-described embodiments are merely illustrative of certain implementations of the present invention, and are described in a relatively specific and detailed manner. However, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements are all within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the appended claims.

Claims

1. A method for intelligent recognition of medical annotation symbols on the body surface and precise positioning of a physiotherapy robot, characterized in that: The method includes the following steps S1. Consultation and annotation stage: Consultants use mobile devices to draw medical annotation symbols on the patient's body surface images and set operation semantics and treatment parameters for each symbol; S2, Image Acquisition and Preprocessing Stage: Acquire color images and depth information of the patient's body surface, and perform contrast enhancement and depth filtering preprocessing. S3, Multi-class symbol detection stage: Real-time target detection is performed on the preprocessed image using a neural network model trained on a specialized dataset, and the symbol category label, bounding box coordinates, and confidence score are output. S4, Geometric Center Point Precise Calculation Stage: Calculate the geometric center pixel coordinates (u*, v*) for each symbol; S5, 3D coordinate mapping stage: Using the camera intrinsic parameter matrix K and the depth value d, the center point pixel coordinates are transformed to the camera coordinate system through the pinhole camera back projection model, and then transformed to the robot base coordinate system through the hand-eye calibration extrinsic parameter matrix T_cam_to_base to obtain millimeter-level 3D physical coordinates; S6. Motion Planning and Execution Phase: The robot generates an ordered sequence of physiotherapy points by parsing and annotating the semantics, performs inverse kinematics solving, fifth-order polynomial trajectory planning, and full-path collision pre-detection. After confirmation by the operator's AR visualization, the robot automatically executes all physiotherapy operations, with real-time collision monitoring and force control feedback throughout the process.

2. The method for intelligent recognition of medical annotation symbols on the body surface and precise positioning of a physiotherapy robot according to claim 1, characterized in that: The neural network model trained on the specific dataset in step S2 is specifically as follows: A specialized dataset of medical annotation symbols for the body surface was constructed and a neural network model was developed to obtain the neural network model. Includes the following steps (1) Raw data collection; (2) Data labeling; (3) Pre-trained model; (4) Model optimization; (5) Verification and evaluation.

3. The method for intelligent recognition of medical annotation symbols on the body surface and precise positioning of a physiotherapy robot according to claim 2, characterized in that: The specific steps (3) are as follows: Pre-trained YOLOv9-C weights on the dataset are used as the starting point for transfer learning, and the transfer learning strategy is as follows: Freeze strategy: During training rounds 1-50 (warm-up phase), freeze the weights of the first 10 layers of the backbone network Gelan-C, and only fine-tune the neck FPN+PAN layer and the detection head, with a learning rate lr0=0.001; Full network fine-tuning: In rounds 51-300, all layers are unfrozen, the entire network participates in gradient updates, and the learning rate decays from 0.01 to 0.0001 using a cosine annealing strategy; Data augmentation configuration: Enable Mosaic enhancement (4 images stitched together, probability = 1.0), random horizontal / vertical flip (probability = 0.5), random rotation (±15°), HSV color gamut perturbation (h = 0.015, s = 0.7, v = 0.4), random noise and blur; Training hyperparameters: batch_size=16, input size 640×640, optimizer SGD (momentum=0.937, weight_decay=0.0005), training on an NVIDIA A100 GPU took approximately 4 hours; The total training loss is calculated using the PGI auxiliary branch weighted joint loss: L_total = L_main + λ × L_aux; In the formula, L_main is the multi-task loss of the main detection head, L_aux is the supervision loss of the PGI auxiliary invertible branch, and λ is the balancing weight (default value is 0.25). Both losses are backpropagated to the shared backbone feature extraction network. The main head loss L_main consists of three parts: L_main = L_cls + L_box + L_dfl L_cls = Σ [-y·log(p) - (1-y)·log(1-p)] (Binary cross-entropy classification loss) L_box = 1 - CIoU(b_pred, b_gt) (Full IoU bounding box regression loss) L_dfl = Σ [-(y_i+1-y)·log(S_i) - (y-y_i)·log(S_{i+1})] (Distribution focus loss) In the formula, L_cls is the binary cross-entropy classification loss, y is the ground truth class label, and p is the predicted probability; L_box uses CIoU (complete IoU) loss, which comprehensively considers the overlap rate, center point distance, and aspect ratio; L_dfl is the distribution focus loss, which models the bounding box coordinates as a discrete probability distribution to improve the fine-grained positioning accuracy; b_pred and b_gt are the predicted box and the ground truth box, respectively; S_i is the Softmax probability output.

4. The method for intelligent recognition of medical annotation symbols on the body surface and precise positioning of a physiotherapy robot according to claim 1, characterized in that: Step S5 specifically includes S51, Camera intrinsic parameter matrix K (pinhole camera model): In the formula, fx and fy are the focal lengths of the camera in the x / y directions (in pixels), and cx and cy are the coordinates of the optical principal point (in pixels), which are precisely determined by the checkerboard calibration program, with a reprojection error of no more than 0.5 pixels; S52, Pixel coordinates to camera 3D coordinates back projection: Xc = (u* - cx) × d / fx Yc = (v* - cy) × d / fy Zc = d In the formula, (u*,v*) are the center point pixel coordinates output in step S3; d is the depth measurement value of the depth sensor at this pixel position (in mm, which needs to be registered with the RGB image for depth-color); (Xc,Yc,Zc) are the three-dimensional coordinates of the center point in the camera coordinate system, in mm; S53. Transformation from camera coordinate system to robot base coordinate system: P_base = T_cam_to_base × [Xc, Yc, Zc, 1]ᵀ In the formula, T_cam_to_base is the 4×4 homogeneous transformation matrix obtained by hand-eye calibration in step S1.2, which includes a 3×3 rotation submatrix R and a 3×1 translation vector t; P_base = [Xw,Yw,Zw,1]ᵀ is the three-dimensional physical coordinate of the center point in the robot base coordinate system, in mm, with an accuracy better than 2mm; S54. Comprehensive Error Analysis: The final three-dimensional positioning error of the system is synthesized from the root mean square of various errors: σ_total = √(σ_detect² + σ_center² + σ_depth² + σ_calib²) In the formula, σ_detect≈1px (YOLOv9 detection box error), σ_center≈0.5px (geometric center algorithm error), σ_depth≈1mm (depth sensor error), and σ_calib≈0.3mm (calibration residual); the synthesized accuracy σ_total < 2mm, which meets the accuracy requirements of clinical physiotherapy.

5. The method for intelligent recognition of medical annotation symbols on the body surface and precise positioning of a physiotherapy robot according to claim 1, characterized in that: Step S6 specifically includes S61, Annotation Semantic Parsing: The system reads the structured annotation task file output in step S1, parses the category semantics of each point and the corresponding operation parameters, and generates an ordered sequence of physiotherapy points; S62. Inverse Kinematics Solution: For each target point P_base in the ordered sequence of therapeutic points, combined with the end-effector approach direction vector n (provided by the symbol category or arrow direction), calculate the robot joint angle target configuration q_tgt: q_tgt = IK(P_base, n, q_seed) In the formula, IK(·) is the robot inverse kinematics solution function (using numerical iteration or analytical method, depending on the robot model); P_base is the three-dimensional coordinates of the target point; n is the end-effector approach direction vector; q_seed is the initial seed joint angle (taking the current joint angle), used to select the optimal configuration in the case of multiple solutions (minimum joint motion criterion). S63. Trajectory Planning and Collision Pre-detection: Before the robot's actual movement, fifth-order polynomial interpolation is used to plan the motion trajectory from the current configuration q_curr to the target configuration q_tgt in joint space, and full-path collision pre-detection is performed. q(t)=a3+a1t+a2t 5 +a3t 5 +a4t 5 +a4t 5 ; In the formula, the coefficients a0, a1, a2, a3, and a4 are uniquely determined by the six boundary conditions of the starting and ending points, velocity (=0), and acceleration (=0), ensuring smooth start and stop of the robot; S64. Visualization and Operator Confirmation: After the trajectory planning is completed, the system displays it in an augmented reality overlay on the operation interface and waits for operator confirmation. S65. Automatic execution and real-time monitoring: The robot automatically visits each physiotherapy point in sequence according to the planned path and performs the corresponding operation actions.

6. A system for intelligent recognition of medical marking symbols on the body surface and precise positioning of a physiotherapy robot, characterized in that: The system is used to perform the method according to any one of claims 1-5, and the system includes... The mobile annotation module allows medical personnel to draw medical annotation symbols on images of patients' bodies, set operational semantics and treatment parameters for each symbol, and generate structured annotation task files. The image acquisition module is used to acquire color images and depth information of the patient's body surface, and to perform contrast enhancement and depth filtering preprocessing on the acquired images; The multi-class symbol detection module uses a neural network model trained on a specialized dataset to perform real-time target detection on preprocessed images and output symbol category labels, bounding box coordinates, and confidence scores. The geometric center accurate calculation module is used to execute category-specific geometric algorithms based on the detected symbol category and bounding box to accurately calculate the pixel coordinates of the geometric center of each symbol; The 3D coordinate mapping module is used to transform the center point pixel coordinates to the camera coordinate system through the pinhole camera back projection model using the camera intrinsic parameter matrix and depth value, and then transform them to the robot base coordinate system through the hand-eye calibration extrinsic parameter matrix to obtain the 3D physical coordinates. The semantic parsing and motion planning module is used to parse the annotation semantics in the structured annotation task file, generate an ordered sequence of physiotherapy points, and perform inverse kinematics solving, trajectory planning, and full-path collision pre-detection. The augmented reality visualization confirmation module is used to overlay and display the detected symbol points, the planned motion path of the robot end effector, and the operating parameters of each physiotherapy point on the operation interface in an augmented reality manner, so that the operator can confirm or adjust the execution plan. The robot execution and safety monitoring module is used to control the robot to automatically perform physiotherapy operations according to the planned path, and to continuously perform real-time collision monitoring and force control feedback during the execution process.

7. The intelligent recognition system for medical annotation symbols on the body surface and the precise positioning system for a physiotherapy robot according to claim 6, characterized in that: The multi-class symbol detection module uses a Gelan-C generalized efficient layer aggregation network as the backbone feature extraction network, and is trained by combining a programmable gradient information-assisted reversible branch mechanism; the programmable gradient information-assisted branch only participates in gradient calculation during the training phase and does not participate in calculation during the inference phase.

8. The intelligent recognition system for medical annotation symbols on the body surface and the precise positioning system for a physiotherapy robot according to claim 6, characterized in that: The precise geometric center calculation module includes: The circle symbol calculation submodule is used to calculate the geometric center of the circle symbol using the least squares circle fitting method; The cross symbol calculation submodule is used to calculate the geometric center of the cross symbol using the double-stroke PCA principal axis intersection method; The triangle symbol calculation submodule is used to calculate the geometric center of a triangle symbol using the centroid method of three vertices; The star symbol calculation submodule is used to calculate the geometric center of the star symbol using the convex hull center method; The square symbol calculation submodule is used to calculate the geometric center of the square symbol using the minimum bounding rectangle center method; The direction arrow symbol calculation submodule is used to calculate the coordinates of the force application point and the direction angle of the direction arrow symbol using the skeleton root positioning and direction angle extraction method.

9. The intelligent recognition system for medical annotation symbols on the body surface and the precise positioning system for a physiotherapy robot according to claim 8, characterized in that: The semantic parsing and motion planning module inputs the three-dimensional coordinates of the cross-type avoidance symbol as an obstacle constraint into the collision pre-detection module. During the trajectory planning stage, it automatically sets a safe distance constraint for each link of the robot for the avoidance point, ensuring that the robot maintains a safe gap of not less than 30mm with the avoidance area during execution. The square symbol calculation submodule is also used to automatically generate a raster scanning path based on the major axis direction, minor axis direction and size of the smallest bounding rectangle. The robot end effector performs parallel reciprocating scanning in rows within the square area, and the row spacing is automatically calculated based on the diameter of the end tool.

10. The intelligent recognition system for medical annotation symbols on the body surface and the precise positioning system for a physiotherapy robot according to claim 8, characterized in that: The mobile annotation module supports gesture drawing mode and symbol toolbar selection mode. The drawn annotation information is transmitted to the semantic parsing and motion planning module in a structured JSON file format. The robot execution and safety monitoring module performs rolling verification of the collision risk of the future predicted position at a frequency of not less than 10Hz and updates the execution progress in real time. The operator can stop the robot's movement at any time through the emergency stop button.

Images