Method and system for generating orthopedic insole model

Abstract

The invention discloses a method and a system for generating an orthopedic insole model, relates to the technical field of foot orthopedic, and aims to solve the problems of low orthopedic insole generation efficiency and low insole matching accuracy in the prior art. The method comprises the following steps: acquiring a sole spatial-temporal feature vector, a foot static parameter and an initial orthopedic insole model of a user; the multi-dimensional feature vectors are input into a double-task machine learning model, and shape righting insole deformation parameters are output; the multi-dimensional feature vector pair is obtained based on splicing of foot static parameters and foot sole space-time feature vectors; and processing the initial orthopedic insole model based on the orthopedic insole deformation parameters to generate a target orthopedic insole model. According to the method, the personalized requirements of the user on the orthopedic insole can be efficiently and accurately met.

Description

Technical Field

[0001] This invention relates to the field of foot orthosis technology, and more particularly to a method and system for generating orthotic insole models. Background Technology

[0002] Orthotic insoles, as an important rehabilitation method for intervening in foot deformities, improving gait function, and relieving foot pain, have received widespread attention in recent years due to their custom technology.

[0003] Traditional orthotic insole manufacturing methods suffer from low precision, long lead times, and reliance on technician experience, making it difficult to meet the demands of large-scale and personalized production. Researchers both domestically and internationally are gradually introducing digital technologies to drive the development of orthotic insole customization towards greater efficiency and precision.

[0004] While existing digital orthotic insole customization solutions have improved efficiency to some extent by introducing scanning and pressure detection technologies, they still have significant limitations. These solutions generally rely on manual selection and assembly of prefabricated modules. The customization process is highly subjective, involves many manual steps, and is time-consuming, resulting in a low matching rate between the orthotic insole and the foot, as well as low generation efficiency. At the same time, they do not adequately integrate dynamic gait characteristics during walking, making it difficult to adapt to the spatiotemporal changes in plantar pressure during human movement. Consequently, the accuracy and personalization of insole matching cannot meet the correction needs of patients with different foot problems, especially the elderly. Summary of the Invention

[0005] The purpose of this invention is to provide a method and system for generating orthotic insole models, so as to more efficiently and accurately meet users' personalized needs for orthotic insoles.

[0006] To achieve the above objectives, the present invention provides the following technical solution: In a first aspect, the present invention provides a method for generating an orthotic insole model, comprising: Obtain the user's spatiotemporal feature vector of the foot, static parameters of the foot, and initial orthotic insole model; The multidimensional feature vector is input into a dual-task machine learning model, and the orthopedic insole deformation parameters are output; the multidimensional feature vector is obtained by concatenating the foot static parameters and the foot spatiotemporal feature vector. Based on the deformation parameters of the orthopedic insole, the initial orthopedic insole model is processed to generate the target orthopedic insole model.

[0007] Optionally, before obtaining the user's plantar spatiotemporal feature vector and the three-dimensional point cloud model of the foot, the method further includes: determining the user's plantar spatiotemporal feature vector; Determine the user's plantar spatiotemporal feature vector, including: A wearable pressure detection system is used to collect dynamic pressure time-series data at a preset number of points on the soles of the user's feet; Extract the pressure time integral and pressure center trajectory spatiotemporal features of each point from the preset number of points from the dynamic pressure time series data; the pressure center trajectory spatiotemporal features include at least the envelope area, trajectory length, and average speed of the pressure center point moving within the corresponding time. The pressure time integral and the spatiotemporal features of the pressure center trajectory are determined as the plantar spatiotemporal feature vector.

[0008] Optionally, before acquiring the user's spatiotemporal feature vector of the foot and the three-dimensional point cloud model of the foot, the method further includes: A three-dimensional point cloud model of the foot is constructed, and the static parameters of the foot and the initial orthotic insole model are determined based on the three-dimensional point cloud model of the foot. Constructing a 3D point cloud model of the foot, including: Full-foot multi-view point cloud data were acquired using optical scanning equipment; Multiple feature points are selected from the full-foot multi-view point cloud data, and a high-dimensional descriptor is generated for each of the multiple feature points to obtain multiple high-dimensional descriptors; the multiple feature points include at least points corresponding to high curvature regions, sharp edges, and corners; For each pair of high-dimensional descriptors among the multiple high-dimensional descriptors, the formula is used:

[0009] The Euclidean distance between each pair of high-dimensional descriptors is calculated. Substituting the Euclidean distance into the formula:

[0010] The set of reliable matching point pairs corresponding to the plurality of high-dimensional descriptors is calculated; wherein, For point clouds The first in A high-dimensional descriptor vector of feature points; For point clouds The first in A high-dimensional descriptor vector of feature points; The similarity threshold; A set of reliably matched point pairs; Align each point cloud data in the set of reliable matching point pairs to a unified coordinate system to generate the three-dimensional point cloud model.

[0011] Optionally, determining the static parameters of the foot based on the three-dimensional point cloud model of the foot includes: Identify the position coordinates of multiple marker points in the three-dimensional point cloud model; the multiple marker points include at least the tip of the longest toe, the lowest point of the first metatarsal head, the most convex point of the first metatarsal head, the highest point of the arch, the lowest point of the fifth metatarsal head, the most convex point of the fifth metatarsal head, the lowest point of the center of the heel, and the far end of the foot. The foot static parameters are calculated based on the position coordinates of the multiple marker points; the foot static parameters include foot length, foot width, and arch height.

[0012] Optionally, determining the initial orthotic insole model based on the three-dimensional point cloud model of the foot includes: Using the lowest point of the heel center, the lowest point of the first metatarsal head, and the lowest point of the fifth metatarsal head in the three-dimensional point cloud model of the foot as references, a reference plane is obtained by fitting. A cutting plane is generated by offsetting upward by a first preset distance along the normal direction of the reference plane; The foot three-dimensional point cloud model is cut using the cutting plane and all point cloud data below the cutting plane are retained to generate the upper surface curved surface of the orthopedic insole model; The upper surface of the orthopedic insole model is defined as the initial orthopedic insole model.

[0013] Optionally, the foot static parameters and the foot spatiotemporal feature vector are input into a dual-task machine learning model to output orthotic insole deformation parameters; The multidimensional feature vector is learned by using the fully connected layer in the dual-task machine learning model to obtain the learned features. The learned features are classified and identified using the AdaBoost model in the dual-task machine learning model to obtain foot type categories; the foot type categories include at least normal foot, flat foot, and high arch foot. The learned features are processed using the regression network model in the dual-task machine learning model to output the orthopedic insole deformation parameters; the orthopedic insole deformation parameters include the deformation height parameter and the deformation influence range parameter of the area to be deformed.

[0014] Optionally, the initial orthotic insole model is processed based on the orthotic insole deformation parameters, including: For the region requiring deformation, a Gaussian decay function is used:

[0015] The upper surface of the initial orthotic insole model is subjected to gradient deformation; wherein, This represents the gradient deformation height at the current point. The deformation height parameter predicted by the regression network model; This is the Euclidean distance from the deformation center point to the current point; The parameter represents the range of deformation influence.

[0016] Optionally, processing the initial orthotic insole model based on the orthotic insole deformation parameters further includes: After deforming the initial orthotic insole model, the upper surface of the deformed insole is triangulated using Delaunay 2.5D, and the bottom surface of the orthotic insole is generated by offsetting downwards by a second preset distance along the normal direction of the reference plane. The deformed upper surface of the insole and the flat bottom surface of the orthopedic insole are combined to generate the target orthopedic insole model.

[0017] Compared with existing technologies, this invention provides a method for generating orthotic insole models. It employs a dual-task machine learning model to process static and dynamic biomechanical features, namely, static foot parameters and spatiotemporal feature vectors of the foot, outputting orthotic insole deformation parameters. Based on these deformation parameters, the initial orthotic insole model is then processed to generate a target orthotic insole model. This eliminates the tedious process of manually measuring and analyzing orthotic insole deformation parameters required in existing technologies. Furthermore, the dual-task machine learning model provides a more intelligent and accurate way to obtain these parameters, reducing time costs and improving insole matching accuracy, thus meeting users' personalized needs for orthotic insoles.

[0018] Secondly, the present invention also provides a system for generating orthotic insole models, comprising: a wearable pressure detection system, an optical scanning device, and a controller; the wearable pressure detection system and the optical scanning device are both connected to the controller; the controller integrates a dual-task machine learning model; The wearable pressure detection system is used to determine the spatiotemporal feature vector of the user's foot. The optical scanning device is used to scan the user's feet to construct a three-dimensional point cloud model of the user's feet; The controller is used to determine the static parameters of the foot and the initial orthotic insole model based on the three-dimensional point cloud model of the foot, and to concatenate the static parameters of the foot and the spatiotemporal feature vector of the foot to obtain a multidimensional feature vector; The dual-task machine learning model is used to process the input multidimensional feature vector and output the orthopedic insole deformation parameters. The controller is also used to process the initial orthopedic insole model based on the orthopedic insole deformation parameters to generate a target orthopedic insole model.

[0019] Optionally, the wearable pressure detection system integrates a distributed flexible sensor array. Attached Figure Description

[0020] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this invention, illustrate exemplary embodiments of the invention and are used to explain the invention, but do not constitute an undue limitation of the invention. In the drawings: Figure 1 This is a flowchart illustrating the method for generating an orthopedic insole model according to an embodiment of the present invention; Figure 2 This is a schematic diagram of the wearable pressure detection system provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the customized human-computer interaction software interface provided in an embodiment of the present invention; Figure 4 This is a schematic diagram of a three-dimensional foot scan using an optical scanning device provided in an embodiment of the present invention; Figure 5 This is a schematic diagram illustrating the state of the construction process of the three-dimensional point cloud model of the foot provided in an embodiment of the present invention; Figure 6 This is a schematic diagram showing the locations of the longest toe tip, the lowest point of the first metatarsal head, the most prominent point of the first metatarsal head, and the highest point of the arch of the foot, provided in an embodiment of the present invention. Figure 7 This is a schematic diagram showing the locations of the lowest point of the fifth metatarsal head, the most prominent point of the fifth metatarsal head, the lowest point of the center of the heel, and the last endpoint, provided in an embodiment of the present invention. Figure 8 This is a schematic diagram of the upper surface of the orthopedic insole generated after cropping a three-dimensional point cloud model of the foot, as provided in an embodiment of the present invention. Figure 9 This is a schematic diagram of the network structure of the AdaBoost-GNN dual-task machine learning model provided in an embodiment of the present invention; Figure 10 This is a schematic diagram of the orthopedic insole surface before gradient deformation using a Gaussian decay function, provided in an embodiment of the present invention. Figure 11 This is a schematic diagram of the orthopedic insole surface after gradient deformation using a Gaussian decay function, provided in an embodiment of the present invention. Figure 12 This is a typical parameter configuration diagram provided in the embodiments of the present invention; Figure 13 This is a comparison diagram of the pressure gradient distribution before deformation and the pressure gradient distribution after deformation provided in the embodiments of the present invention; Figure 14 This is a schematic diagram of performing Delaunay 2.5D triangulation on the deformed insole surface and generating a solid model according to an embodiment of the present invention; Figure 15This is a schematic diagram of the structure of the orthopedic insole model generation system provided in this embodiment of the invention. Detailed Implementation

[0021] To facilitate a clear description of the technical solutions in the embodiments of the present invention, the terms "first" and "second" are used to distinguish identical or similar items with essentially the same function and effect. For example, the first threshold and the second threshold are merely used to distinguish different thresholds and do not limit the order of execution. Those skilled in the art will understand that the terms "first" and "second" do not limit the quantity or execution order, and that the terms "first" and "second" are not necessarily different.

[0022] It should be noted that in this invention, the terms "exemplary" or "for example" are used to indicate examples, illustrations, or descriptions. Any embodiment or design described as "exemplary" or "for example" in this invention should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of terms such as "exemplary" or "for example" is intended to present the relevant concepts in a concrete manner.

[0023] In this invention, "at least one" means one or more, and "more than one" means two or more. "And / or" describes the relationship between the associated objects, indicating that three relationships can exist.

[0024] like Figure 1 As shown, this embodiment of the invention provides a method for generating an orthotic insole model, which may include: Step 100: Obtain the user's spatiotemporal feature vector of the foot, static parameters of the foot, and initial orthotic insole model.

[0025] Before proceeding to step 100, it is necessary to predetermine the spatiotemporal feature vector of the user's foot and to construct a three-dimensional point cloud model of the user's foot in advance.

[0026] In one optional embodiment, determining the user's plantar spatiotemporal feature vector may specifically include: Step 1: Use a wearable pressure detection system to collect dynamic pressure time-series data at a preset number of points on the soles of the user's feet.

[0027] For example Figure 2 As shown, the wearable pressure detection system may include a distributed flexible sensor array and a main control box.

[0028] For example, see Figure 2 The distributed flexible sensor array can employ the RX-ES42-18 array-type flexible thin-film pressure sensor. This array uses a 6-row, 4-column layout with a total of 18 independent sensing points, evenly covering the key areas of the sole of the foot. The single-point measurement range is 70kg, sufficient to cover the pressure range of the human foot.

[0029] See Figure 2 The distributed flexible sensor array is connected to the main control box via flexible wires. The main control box integrates three key modules: a signal conditioning circuit, a microprocessor, and a Bluetooth module. The signal conditioning circuit receives the raw analog electrical signals from the flexible sensor array, amplifies and filters them, providing a high-quality, stable analog signal for subsequent analog-to-digital conversion. The microprocessor's core function is control and data processing; it controls the analog-to-digital converter to convert the conditioned analog signal into a digital signal and periodically reads digital pressure data at a preset 50Hz sampling frequency. The Bluetooth module establishes a wireless communication link with the host computer, wirelessly transmitting the digital pressure data sent by the microprocessor to the software on the host computer.

[0030] As for the software in the host computer, customized human-computer interaction software can be developed using the Python language in the PyCharm integrated development environment, such as... Figure 3 As shown, the main functions of the software include: Data reception: Bluetooth serial communication is established through the PySerial library to continuously receive pressure data from 18 points on the main control box at a sampling interval of 20ms. The software interface provides baud rate settings, typically 9600 or 115200, device scanning, and connection / disconnection operation buttons.

[0031] Data processing: The software incorporates a real-time algorithm for calculating the center of pressure (COP) coordinates. After each sampling, the formula is used: (1) Calculate the coordinates of the pressure center; where, The x-coordinate of the pressure center point The ordinate of the pressure center point. For the first i The x-coordinate of each sensor, For the first i The ordinate of each sensor, For the first i Pressure sampling values ​​from each sensor.

[0032] The software calculates the COP coordinates at the current moment using the coordinates of 18 points and real-time pressure values. It provides "Start" and "Stop" buttons for recording the COP trajectory over a period of time.

[0033] Image observation: The software interface draws a silhouette of the foot and marks the 18 sensor points on the map according to their physical layout. The pressure values ​​of each point are displayed intuitively through color mapping, generating a dynamic heat map.

[0034] For example, Figure 3The key parameters of the customized human-computer interaction software are detailed below: The software interface displays the title "Pressure Monitoring Terminal". On the left, there are serial port connection settings, the serial port number is selectable, the baud rate is selectable between 9600bps and 115200bps, and the default is 115200bps to balance transmission stability and speed. There are also "Connect" and "Disconnect" buttons. Below is the serial port data display area. Next is the COP coordinate section, which is divided into the horizontal coordinate X and vertical coordinate Y of the left and right feet, as well as "Start" and "Stop" buttons. The data generation area includes input boxes for name, gender, weight, shoe size, and arch height. Foot type can be selected as "normal", and the data type can be selected as "static" or "dynamic". There are also "Start Generation" and "Stop" buttons related to ".csv file". On the right is the foot pressure (unit: g) display area, which shows the outline of the left and right feet and the pressure distribution. The data sampling interval is fixed at 20ms, which works in conjunction with the 50Hz sampling frequency of the main control box to ensure complete acquisition of pressure data. The pressure center (COP) calculation parameters include preset coordinate matrices of multiple sensors, with the horizontal axis range set to 0-140mm and the vertical axis range set to 0-220mm, matching the typical size range of an adult foot. The pressure display uses a specific mapping scheme to map pressure values ​​to different visual effects. At the same time, a warning threshold can be set. When the pressure at a certain point exceeds the threshold, the software will automatically mark the area and issue a visual prompt. The time range for COP trajectory recording can be customized within 10-60 seconds. The default is to record 30 seconds of complete gait cycle data, which is convenient for extracting stable spatiotemporal features.

[0035] Step 2: Extract the pressure time integral and pressure center (COP) trajectory spatiotemporal features of each point from the dynamic pressure time series data, which are a preset number of points. The pressure center trajectory spatiotemporal features include at least the envelope area, trajectory length, and average velocity of the pressure center point moving within the corresponding time period. For example, the preset number is 18.

[0036] In specific implementations, for example, the pressure-time integral values ​​of 18 points over the entire cycle are extracted from dynamic pressure time series data to generate an 18-dimensional feature vector. Furthermore, the envelope area, total trajectory length, and average velocity of the COP trajectory within that cycle are calculated to form a 3-dimensional spatiotemporal feature vector.

[0037] Step 3: Determine the pressure time integral and the spatiotemporal characteristics of the pressure center trajectory as the plantar spatiotemporal feature vector.

[0038] For example, the 18-dimensional feature vector and the 3-dimensional spatiotemporal feature vector are concatenated to generate a 21-dimensional dynamic biomechanical feature vector, which is the aforementioned plantar spatiotemporal feature vector.

[0039] The beneficial effects of this embodiment: 1) From a hardware perspective, wearable stress detection systems ensure the accuracy and completeness of dynamic data.

[0040] Employing the RX-ES42-18 array-type flexible thin-film pressure sensor, it features an 18-point layout with 6 rows and 4 columns. This layout can evenly cover key areas of the sole of the foot, such as the heel, metatarsals, and arch. Furthermore, the 70kg range per point covers the pressure range of the human foot, accurately capturing dynamic pressure changes in each key area and avoiding feature omissions due to incomplete sensor coverage.

[0041] The main control box integrates a signal conditioning circuit, a microprocessor, and a Bluetooth module. The signal conditioning circuit amplifies and filters the raw analog signal to reduce environmental interference, such as noise caused by slight sensor displacement during walking, ensuring the authenticity of the pressure data. The 50Hz sampling frequency corresponds to a 20ms sampling interval, which is transmitted wirelessly via Bluetooth, avoiding interference from wired connections on the user's gait. The collected dynamic pressure time-series data more closely resembles the actual walking state, providing a high-quality data source for subsequent extraction of pressure time integrals and COP trajectory spatiotemporal features.

[0042] 2) From a software perspective, software functionality improves process automation and data readability.

[0043] The data receiving module supports compatibility with different hardware devices, lowering the barrier to entry for system use. At the same time, it receives data in real time at 20ms intervals, ensuring the continuity of dynamic data and avoiding COP trajectory breakage caused by sampling interruption.

[0044] COP trajectory is a core indicator of gait stability. The software's built-in algorithm automatically calculates COP coordinates, replacing the traditional method of manually marking trajectories, reducing human error. The start-stop recording function can collect complete gait cycle data, such as 1-2 walking cycles. The extracted envelope area, trajectory length, and average speed can accurately reflect the degree of gait deviation of the user. For example, the COP trajectory is more dispersed when the gait is unstable in the elderly.

[0045] The pressure heatmap visualization function can convert pressure values ​​into gradient colors to intuitively present high-pressure areas on the sole of the foot. This not only helps medical staff quickly determine foot problems, but also provides visual assistance for subsequent dual-task models to identify foot type and output deformation parameters. For example, the heatmap can directly confirm abnormal pressure in the arch area of ​​patients with foot problems.

[0046] 3) It is especially suitable for the clinical needs of special populations.

[0047] For the elderly, flexible sensors fit the sole of the foot and transmit wirelessly, avoiding deliberate changes in gait due to hardware discomfort. The collected dynamic data can better reflect the actual gait stability of the elderly, and the insoles generated based on this data can more accurately correct gait deviations.

[0048] For users with sports injuries, the 18-point sensor can capture changes in pressure distribution at the injured site. Pressure-time integration can quantify the cumulative force at the injured site, providing a basis for force control in the design of insole deformation parameters, such as reducing pressure at the injured site and avoiding secondary injuries.

[0049] In one alternative embodiment, constructing a 3D point cloud model of the user's feet may specifically include: Step 1: Use an optical scanning device to acquire full-foot multi-view point cloud data.

[0050] In practice, users are required to keep their feet bare and suspended in the air. A structured light 3D scanner, such as EXScanPro, is used to perform multi-angle scans of the feet by changing the relative angle between the device and the feet. Figure 4 As shown, multiple sets of local point cloud data covering the entire surface of the foot are obtained, which are also known as multi-view point cloud data of the entire foot.

[0051] Step 2: Select multiple feature points from the full-foot multi-view point cloud data, and generate a high-dimensional descriptor for each feature point, resulting in multiple high-dimensional descriptors. See [link to relevant documentation]. Figure 5 .

[0052] For example, multiple feature points include at least the points corresponding to high curvature regions, sharp edges, and corners.

[0053] In practice, before selecting feature points, the full-foot multi-view point cloud data needs to be imported into the corresponding processing software in the controller.

[0054] The high-dimensional descriptor is generated using the SIFT (Scale Invariant Feature Transform) algorithm, specifically with 128 dimensions. The generation process includes the following steps: Step 1: Scale space construction is performed on the full-foot multi-view point cloud data. A Gaussian filter with σ=1.6 is used to perform multi-scale convolution on the original point cloud grayscale image (generated by point cloud coordinate mapping) to obtain image pyramids of different scales. Step 2: Detect extreme points in the image pyramid by comparing each pixel with its 8 neighboring pixels and 9 pixels at the adjacent scales above and below to determine the location and scale of the feature points; Step 3: Calculate the principal direction of the feature point: Using the feature point as the center, calculate the gradient direction histogram in a 16×16 neighborhood (gradient direction range 0-360°, each 10° interval). Take the direction corresponding to the peak value in the histogram as the principal direction of the feature point. If there are multiple peak values ​​(the difference is less than 80% of the peak value), retain multiple principal directions to enhance rotation robustness. Step 4: Generate a 128-dimensional descriptor: Divide the neighborhood of the feature point into 4×4 sub-regions, and calculate the gradient histogram of each sub-region in 8 directions to obtain a 4×4×8=128-dimensional vector. Normalize the vector (to eliminate the influence of illumination) to finally obtain a 128-dimensional high-dimensional descriptor vector.

[0055] See Figure 5 , A high-dimensional descriptor vector representing a certain feature point, where, The spatial coordinate parameters of the feature point are used to accurately locate the position of the feature point in three-dimensional space; , The curvature temporal parameter of the feature point is used to quantify the dynamic characteristics of the curvature of the region as a function of factors such as viewing angle. This is a high-dimensional descriptor vector of the feature point from another perspective. Spatial coordinate parameters of feature points ； ， These are the corresponding curvature timing parameters; and The system identifies the same feature point from different perspectives; the curvature distribution describes the distribution of curvature in the area surrounding the feature point, helping to determine the characteristics of the feature point; the normal vector determines the direction of the local plane where the feature point is located, and helps to establish the spatial orientation matching relationship of the feature point from different perspectives during feature matching and other processes, so as to achieve more accurate feature matching and finally complete the reconstruction of the three-dimensional point cloud model of the foot.

[0056] Step 3: For pairwise high-dimensional descriptors among multiple high-dimensional descriptors, use the formula: (2) Calculate the Euclidean distance between pairwise high-dimensional descriptors; Step 4: Substitute the Euclidean distance into the formula: (3) A set of reliable matching point pairs corresponding to multiple high-dimensional descriptors is calculated; among them, For point clouds The first in A high-dimensional descriptor vector of feature points; For point clouds The first in A high-dimensional descriptor vector of feature points; The similarity threshold; A set of reliably matched point pairs; For point clouds The High-dimensional descriptor vectors and point clouds of feature points The The Euclidean distance between the high-dimensional descriptor vectors of feature points.

[0057] Step 5: Align each point cloud data in the set of reliable matching point pairs to a unified coordinate system to generate a 3D point cloud model of the foot. Figure 5 The reconstructed 3D point cloud model.

[0058] Specifically, the point cloud data of each point cloud in the set of reliably matched point pairs is aligned to a unified coordinate system by estimating the rigid body transformation matrix.

[0059] After constructing a 3D point cloud model of the user's foot, the static parameters of the foot and the initial orthotic insole model are determined based on the 3D point cloud model of the foot.

[0060] Static parameters of the foot are determined based on a 3D point cloud model of the foot, including: Step 1: Identify the position coordinates of multiple marker points in the 3D point cloud model; these marker points must include at least the tip of the longest toe (see...). Figure 6 The lowest point of the first metatarsal head (see...) Figure 6 The most prominent point of the first metatarsal head (see...) Figure 6 ), the highest point of the arch (see) Figure 6 The lowest point of the fifth metatarsal head (see...) Figure 7 The most prominent point of the fifth metatarsal head (see...) Figure 7 ), the lowest point of the center of the heel (see Figure 7 ) and the final endpoint (see Figure 7 ).

[0061] For example, landmarks can be anatomical landmarks.

[0062] For example, machine learning methods can be used to automatically identify the aforementioned marker locations.

[0063] Step 2: Calculate the static parameters of the foot based on the coordinates of multiple marker points; the static parameters of the foot include foot length, foot width, and arch height.

[0064] For example, foot length can be the axial distance from the center of the heel to the end of the longest toe.

[0065] For example, foot width can be the axial distance from the most prominent point of the first metatarsal head to the most prominent point of the fifth metatarsal head.

[0066] For example, the height of the arch can be the vertical distance from the highest point of the arch to the lowest point of the first metatarsal head and the lowest point of the center of the heel.

[0067] The initial orthotic insole model was determined based on a 3D point cloud model of the foot, including: Using the lowest point of the heel center, the lowest point of the first metatarsal head, and the lowest point of the fifth metatarsal head in the three-dimensional point cloud model of the foot as references, a reference plane is obtained by fitting. Along the normal direction of the reference plane, an upward offset of a first preset distance is made to generate a cutting plane; for example, the preset distance can be taken in the range of 3-5mm. The 3D point cloud model of the foot is cut using a cutting plane, retaining all point cloud data below the cutting plane, to generate the upper surface of the orthopedic insole model. (See [link]). Figure 8 ; The upper surface of the orthotic insole model is defined as the initial orthotic insole model.

[0068] This embodiment achieves static data modeling through structured light scanning, feature matching, marker point recognition, and cutting plane setting. The technical effects are as follows: 1) 3D point cloud reconstruction ensures the integrity and accuracy of the model.

[0069] Multi-view scanning extracts feature points such as high curvature areas and sharp edges to generate high-dimensional descriptors. Then, similarity is calculated using Euclidean distance to ensure accurate alignment of point clouds from different perspectives. For example, point clouds in complex curved areas such as the heel and arch of the foot are not misaligned. Finally, a complete and seamless 3D point cloud model of the foot is generated, avoiding model loss caused by traditional single-view scanning, such as the blind spot in the inner arch of the foot.

[0070] Aligning with a unified coordinate system ensures that point clouds from multiple perspectives are fused in the same spatial coordinate system, avoiding errors in foot length and width calculations caused by coordinate deviations. For example, the traditional manual measurement of foot length has an error of ±5mm, while this method can control the error within ±1mm, thus improving the accuracy of static parameter calculations.

[0071] 2) Standardization is achieved for automatic identification of marker points and calculation of static parameters.

[0072] The method of automatically identifying the first metatarsal head, fifth metatarsal head, heel center, and highest point of the arch replaces the traditional method of manually touching and locating landmarks, such as technicians manually pressing to find the location of the metatarsal heads. This reduces the positioning error of landmarks caused by differences in human experience, such as the thick fat on the feet of elderly patients, which makes manual positioning prone to errors.

[0073] The system automatically calculates foot length, foot width, and arch height based on the coordinates of landmark points. The output parameters are standardized, such as the foot length being the distance from the center of the heel to the end of the longest toe, avoiding the subjective bias of traditional manual measurement. At the same time, the three core parameters fully reflect the characteristics of the foot type, such as the arch height of patients with high arch foot problems being significantly higher than that of normal feet.

[0074] 3) The initial insole model generation ensures a good fit and reasonable subsequent deformation.

[0075] The reference plane is fitted to the heel center and the lowest point of the first / fifth metatarsal heads to fit the foot support surface, ensuring that the support base of the initial model conforms to the physiological structure of the human foot and avoiding misalignment of the insole support points due to the offset of the reference plane.

[0076] The reference plane is moved up 3-5mm to generate the cutting plane. The offset of 3-5mm is an optimized value that balances fit and support. If the offset is too small, the initial model will be too thin, resulting in insufficient space for subsequent deformation. If the offset is too large, the initial model will be too thick, which may affect wearing comfort. The point cloud below the cutting plane is retained to generate the initial curved surface, ensuring that the initial model fits the foot shape 1:1. Subsequently, only local gradient deformation based on biomechanical characteristics is needed, such as raising the arch area, to reduce the instability of the model structure caused by overall deformation.

[0077] Step 200: Input the multidimensional feature vector into the dual-task machine learning model and output the orthotic insole deformation parameters; the multidimensional feature vector is obtained by concatenating the foot static parameters and the plantar spatiotemporal feature vector.

[0078] For example, see Figure 9 A dual-task machine learning model can be the AdaBoost-GNN dual-task model.

[0079] Step 300: Based on the deformation parameters of the orthotic insole, perform deformation processing on the initial orthotic insole model to generate the target orthotic insole model.

[0080] As can be seen from the above, the method for generating the orthopedic insole model provided in this embodiment of the invention has the following technical effects: 1) In this embodiment, the spatiotemporal feature vector of the user's foot and the static parameters of the foot are obtained to achieve a comprehensive capture of the dynamic and static biomechanical characteristics of the user's foot, replacing the standardized template of the traditional modular solution and ensuring personalization from the data source.

[0081] The dual-task machine learning model outputs orthotic insole deformation parameters. The algorithm automatically completes foot type classification and deformation parameter calculation, completely eliminating the reliance on manual module selection and adjustment in traditional solutions, and realizing the automation of design parameter generation.

[0082] In this embodiment of the invention, the static parameters of the foot and the spatiotemporal feature vector of the sole are concatenated into a multidimensional feature vector, which allows the dual-task machine learning model to directly output deformation parameters based on biomechanical features, establish an automatic mapping relationship between input features and output parameters, without the need for manual intervention, and improve the efficiency of automated generation of orthopedic insoles.

[0083] 2) To address the shortcomings of low matching degree of orthopedic insoles in existing technologies.

[0084] The spatiotemporal feature vector of the foot in this invention specifically incorporates dynamic gait data during walking, which overcomes the limitations of traditional solutions that only focus on static morphology. This allows the final target insole to adapt to changes in foot pressure during human movement, improving gait stability, and especially enhancing the gait stability correction effect for the elderly.

[0085] In an optional embodiment, step 200: inputting the foot static parameters and the plantar spatiotemporal feature vector into a dual-task machine learning model, and outputting orthotic insole deformation parameters, specifically includes: The fully connected layer in the dual-task machine learning model is used to learn features from the multi-dimensional feature vector to obtain the learned features. The AdaBoost model, a dual-task machine learning model, is used to classify and identify the learned features to obtain foot type categories; the foot type categories include at least normal foot, flat foot, and high arch foot. The learned features are processed using a regression network model in a dual-task machine learning model to output orthotic insole deformation parameters. The orthotic insole deformation parameters include the deformation height parameter and the deformation influence range parameter of the area to be deformed (corresponding to the abnormal pressure point or anatomical location).

[0086] For example, combined Figure 9 The dual-task machine learning model is the AdaBoost-GNN dual-task machine learning model. Its structure is a progressive design from input to shared feature learning and then to dual-branch output. The data processing flow and functions of each module are as follows: (1) The overall structure of the model.

[0087] Input layer: Receives a 24-dimensional multidimensional feature vector composed of static foot parameters (foot length, foot width, and arch height) and spatiotemporal feature vectors of the foot, which serves as the raw data basis for model calculation.

[0088] Shared Feature Learning Layer: The core consists of 2-3 fully connected layers that use the ReLU activation function to perform nonlinear transformation and dimensionality optimization on the input 24-dimensional feature vector. This layer does not distinguish between subsequent tasks and focuses on extracting common biomechanical key information from dynamic and static features, such as the correlation between arch height and pressure integral, and the matching features between foot length and COP trajectory length, generating 128-dimensional or 256-dimensional learned features to provide high-quality feature support for the dual-branch task.

[0089] Dual-task output branches: Based on the output of the shared feature learning layer, the output is divided into two parallel and independent task branches, which respectively realize foot type classification and deformation parameter regression.

[0090] AdaBoost classification branch: It consists of N weak classifiers (preferably CART decision trees) connected in series using the AdaBoost ensemble algorithm. Each weak classifier takes the learned features as input and outputs a preliminary foot type prediction result: the prediction result is normal foot, flat foot, or high arch foot.

[0091] Training Phase: During the initial training, all training samples are assigned the same weights, for example, initially set to D1. The first weak classifier is then trained, and its classification error is calculated. Weight coefficients are then assigned based on the error. In subsequent training, the weights of samples misclassified in the previous round are increased (updated to sample weights D2, D3, ..., Dn), and the 2nd to nth weak classifiers are trained sequentially and assigned corresponding weight coefficients. … Finally, the prediction results of n weak classifiers are fused through a weighted voting combination strategy to output the final foot type category label.

[0092] Prediction Phase: The learned features of the new user are input, and after being predicted sequentially by n weak classifiers, the predictions are made according to the weight coefficients determined in the training phase. … The weighted calculation is used to obtain the foot type with the highest probability, such as flat feet with a probability of 0.85 and normal feet with a probability of 0.15, and then flat feet is output.

[0093] GNN Regression Branch: Based on the Graph Neural Network (GNN), the 18 pressure sensing points on the sole of the foot are regarded as nodes of the graph. The node features are the pressure integral of the corresponding point in the learned features, the static parameter correlation value, such as the mapping value between the arch height and the arch area node. The edge construction is based on the anatomical correlation between nodes (for example, the first metatarsal head node is connected to the arch node, and the heel node is connected to the fifth metatarsal head node).

[0094] Computation Phase: A message-passing mechanism allows each node to exchange feature information with its neighbors, learning the biomechanical dependencies between different areas of the foot, such as the force relationship between the arch and heel nodes in patients with flat feet. Finally, through the linear activation function of the output layer, two sets of core deformation parameters are output for key areas of the insole, such as the arch support area and metatarsal decompression area: deformation height parameter. (Unit: mm) and deformation influence range parameters (Unit: mm)

[0095] Based on the structure and prediction logic of the AdaBoost-GNN dual-task machine learning model described above, the following supplements the specific training process of the model to ensure that the model performance meets the parameter output requirements of personalized orthotic insoles.

[0096] The training process for a dual-task machine learning model includes the following: (a) Sources of training data.

[0097] Data collection subjects: 300 subjects with different foot conditions were selected. These subjects included 100 with normal feet, 100 with flat feet, and 100 with high arches. The age range was 18-65 years old, with the elderly (≥60 years old) accounting for 30%. The subjects also included different weights, such as 45-100kg, and shoe sizes, such as 35-46, to ensure data diversity.

[0098] Static data: A three-dimensional point cloud model of each subject's foot was acquired using an optical scanning device, and foot length, foot width, and arch height were calculated as static parameters of the foot.

[0099] Dynamic data: Sampling frequency 50Hz, acquisition time 30 seconds. Dynamic foot pressure time-series data was collected for each subject during walking using a wearable pressure detection system. Pressure-time integrals and COP trajectory spatiotemporal features at 18 points were extracted as the foot spatiotemporal feature vector; the COP trajectory spatiotemporal feature envelope area is 0-1000mm². 2 The trajectory length is 0-500mm and the average speed is 0.5-3mm / s.

[0100] Label data: Based on the subjects' foot X-rays and gait analysis results, three senior foot and ankle physicians jointly labeled the foot type: normal foot, flat foot, and high arch foot. At the same time, based on the abnormal areas of plantar pressure distribution, such as the pressure concentration area of ​​the arch of flat feet and the pressure concentration area of ​​the heel and metatarsals of high arch feet, the deformation parameters of the orthotic insoles were manually labeled: deformation height H: 1-8mm, deformation influence range σ: 8-32mm, as label data for the regression task.

[0101] Data preprocessing: The collected static parameters and dynamic feature vectors are normalized and abnormal data are removed, such as pressure values ​​exceeding the sensor range and samples with inconsistent foot type labels. Finally, 2800 sets of valid training samples are obtained, of which 200 sets are used as test samples.

[0102] (II) Training steps.

[0103] Model structure initialization: The dual-task machine learning model includes a shared feature layer, an AdaBoost classification branch, and a regression network branch; Shared feature layer: Two fully connected layers are set. The first layer has an input dimension of 24: 3D foot static parameters and 21D plantar spatiotemporal feature vector, and an output dimension of 64. The activation function is ReLU. The second layer has an input dimension of 64 and an output dimension of 128. The activation function is ReLU.

[0104] AdaBoost classification branch: Select CART decision tree as weak classifier, set the number of weak classifiers to 10, and distribute the initial sample weights evenly: each sample weight = 1 / 280.

[0105] Regression network branch: A 3-layer fully connected network is used. The first layer has an input dimension of 128 and an output dimension of 64, with the activation function ReLU; the second layer has an input dimension of 64 and an output dimension of 32, with the activation function ReLU; the third layer has an input dimension of 32 and an output dimension of 2, corresponding to the deformation height H and the deformation influence range σ, respectively, with the activation function Linear.

[0106] Shared feature layer pre-training: The shared feature layer is trained using static parameters and dynamic feature vectors as inputs, foot type category labels as supervision signals, and cross-entropy loss function. The optimizer is Adam (learning rate 0.001, weight decay 0.0001). The training is conducted for 50 epochs. After each epoch, the accuracy is evaluated on the validation set. When the validation accuracy does not improve for 5 consecutive epochs, the pre-training is stopped and the shared feature layer parameters are saved.

[0107] AdaBoost classification branch training: First round of training: Using the output of the pre-trained shared feature layer as input, train the first CART weak classifier, calculate the classification error (error = sum of weights of incorrect samples). If the error > 0.5, reinitialize the weak classifier; otherwise, calculate the weights of the weak classifier. .

[0108] Sample weight update: For correctly classified samples, the weights are updated to... (Z is the normalization factor); for misclassified samples, the weights are updated to... .

[0109] Repeated training: Train the next 9 weak classifiers according to the above steps. Each weak classifier is trained based on the updated sample weights. Finally, the foot type classification result is obtained by weighted voting.

[0110] Regression network branch training: The parameters of the shared feature layer are fixed, the output of the shared feature layer is used as input, the deformation parameter label (H, σ) is used as the supervision signal, the mean squared error loss function is used to train the regression network, the optimizer is Adam (learning rate 0.0005, weight decay 0.0001), the training is carried out for 80 rounds, and the mean absolute error (MAE) is calculated on the validation set after each round. When the MAE does not decrease for 5 consecutive rounds, the training is stopped.

[0111] Joint model fine-tuning: Unfreeze the shared feature layer and use Loss_total=0.4×Loss_cls+0.6×Loss_reg as the total loss function to jointly fine-tune the entire dual-task model. The training runs for 30 rounds to further improve the model's classification and regression accuracy.

[0112] (iii) Performance evaluation indicators.

[0113] Foot type classification task evaluation metrics: Accuracy: Number of correctly classified samples / total number of samples, target value ≥ 92%; Precision: Number of correct predictions for each category (normal foot / flat foot / high arch) / Total number of predictions for that category, target value ≥ 90%; Recall: Correct predictions for each category / Total true counts for that category; Target value ≥ 90%; Other indicators can be found in relevant technical documents, and will not be elaborated here.

[0114] Analysis of the beneficial effects of this embodiment: 1) Completely realize automated feature-to-parameter mapping, eliminating reliance on manual intervention.

[0115] Compared to traditional methods that require manual selection of prefabricated modules and manual parameter adjustment, the AdaBoost-GNN dual-task machine learning model in this embodiment, through shared feature learning and automatic dual-branch calculation, can complete the entire process from dynamic and static biomechanical features to foot type category and deformation parameters without any manual intervention. For example, after inputting features from elderly patients with flat feet, the model can automatically output the "flat feet" category and deformation parameters of "arch H=4mm, σ=16mm", reducing parameter generation time from the traditional 1-2 hours to minutes or even seconds, significantly improving customization efficiency.

[0116] 2) Improve the accuracy of foot type classification and provide category guidance for deformation parameters.

[0117] The AdaBoost ensemble algorithm addresses the misclassification of complex foot types (such as mild flatfoot and normal foot boundary cases) by weighted fusion of multiple weak classifiers, such as traditional SVMs. Validated with 1000 sets of clinical data, this branch achieves a foot type classification accuracy of over 90%, significantly higher than the approximately 70% accuracy of manual judgment. Furthermore, the classification results indirectly influence the GNN regression branch through shared feature layers, ensuring a high degree of matching between deformation parameters and foot type categories.

[0118] 3) Adapt to the biomechanical correlation of the foot to avoid stress concentration caused by local deformation.

[0119] The GNN regression branch, through its graph structure design of nodes and edges, naturally aligns with the biomechanical relationships of different areas of the foot, such as how changes in heel pressure affect metatarsal pressure distribution. Compared to traditional isolated area deformation calculations, the deformation parameters output by this branch consider the adaptability of adjacent areas: for example, when calculating the deformation of the arch decompression zone for patients with high arches, the σ value of the heel support zone is simultaneously adjusted (increased from 12mm to 18mm) to ensure uniform pressure distribution across the insole after deformation; combined with the subsequent Gaussian decay function, it achieves a significant deformation center and a smooth transition around the periphery, effectively avoiding discomfort caused by localized stress concentration, such as metatarsal chafing.

[0120] 4) Support the personalized needs of special populations and enhance clinical applicability.

[0121] Addressing the characteristics of poor gait stability and thin foot fat layer in the elderly, the AdaBoost branch can accurately identify age-related degenerative flat feet (distinguishing them from juvenile flat feet), and the GNN branch outputs deformation parameters with "low height and wide range": for example... , This ensures effective arch support while preventing foot fatigue caused by excessive deformation.

[0122] 5) It is scalable and can adapt to more foot problem scenarios.

[0123] The number of weak classifiers n, GNN nodes and edges in the dual-task machine learning model can be flexibly adjusted according to the addition of foot problems such as clubfoot and hallux valgus. For example, when adding a hallux valgus scenario, a "first metatarsal valgus node" can be added to the GNN and a "hallux valgus foot" classification category can be added to AdaBoost. The new scenario can be adapted without reconstructing the overall structure of the model, thereby improving the long-term application value of the system.

[0124] In an optional embodiment, step 300, which involves deforming the initial orthotic insole model based on orthotic insole deformation parameters, includes: For the region requiring deformation, a Gaussian decay function is used: (4) The upper surface of the initial orthotic insole model undergoes gradient deformation; among which, This represents the gradient deformation height at the current point. The deformation height parameter predicted by the regression network model; This is the Euclidean distance from the deformation center point to the current point; The parameter represents the range of deformation influence.

[0125] It should be noted that using a Gaussian decay function for deformation ensures the most significant effect at the deformation center, and smoothly decays to zero with increasing distance, effectively avoiding stress concentration. Parameters The effect on deformation is crucial: smaller Values ​​that produce localized, steep deformations are suitable for areas requiring precise, concentrated support; larger values... This value produces a smooth, widespread deformation, making it suitable for areas requiring large-area pressure redistribution. The value is automatically determined by the model based on biomechanical requirements, thus ensuring a precise match between the deformation effect and the correction target. This process is performed sequentially or superimposed on all deformation areas, ultimately resulting in a personalized insole upper surface with a smooth transition driven by biomechanics.

[0126] In practical implementation, when using a Gaussian decay function to perform gradient deformation on the target area of ​​the initial orthotic insole model, the deformation effect can be clearly observed through morphological comparison of local areas, for example... Figure 10 , Figure 11 As shown. Figure 10 The surface is in the local curved state before deformation. Its surface retains the basic shape of the initial model and fits the sole of the foot, but it has not been optimized for areas with abnormal pressure. Figure 11 As shown in the deformed local surface, the bulge effect is most significant in the deformation center area (corresponding to the abnormal pressure point on the sole of the foot or the anatomical position that needs support), and gradually and smoothly decays towards the surrounding area without obvious sharp edges or abrupt changes. This perfectly matches the gradient characteristics of the Gaussian decay function, which is strong at the center and weak at the edges, effectively avoiding local stress concentration and laying the morphological foundation for subsequent pressure distribution optimization.

[0127] The deformation parameters output by the dual-task machine learning model directly determine the deformation effect. Figure 12 This section showcases a typical parameter configuration example: Deformation Height Parameter , representing the maximum bulge height at the deformation center; deformation influence range parameter This determines the effective coverage area of ​​the deformation. In the figure, d1=0.5 d2= d3=2 The markings clearly show the spatial pattern of deformation attenuation: within the d1 range (8mm), the deformation height is close to the maximum value. This ensures the support strength of the core area; within the d1 to d2 range (8-16mm), the deformation height rapidly decreases according to a Gaussian function; within the d2 to d3 range (16-32mm), the deformation height gradually approaches zero, achieving seamless connection with the surrounding non-deformable areas. Furthermore, height markings of h1=3.53mm, h2=2.42mm, and h3=0.54mm further quantify the degree of deformation at different distances, verifying the relatively small deformation. Steep deformation, large The parameter characteristics that produce gradual deformation provide an intuitive reference for parameter configuration for different correction needs, such as precise support and large-area pressure.

[0128] One of the core objectives of deformation treatment is to optimize the distribution of plantar pressure gradient. Figure 13 By comparing the changes in pressure gradient before and after deformation, the technical effect is visually presented. In the left image, the pressure gradient distribution before deformation is uneven, with obvious areas of concentrated high pressure, such as the metatarsals or abnormal arch areas, which can easily lead to excessive local stress and pain during walking. In the right image, the pressure gradient distribution after deformation shows a uniform diffusion trend, with a significant reduction in pressure values ​​in the original high-pressure areas, and the pressure transmission path is more in line with human biomechanical principles. This change fully demonstrates that gradient deformation achieved through a Gaussian decay function not only optimizes the physical shape of the insole but also fundamentally improves the transmission characteristics of plantar pressure, especially catering to the needs of the elderly or patients with foot deformities for gait stability and pressure dispersion, ultimately achieving a technical closed loop of "shape adaptation → pressure optimization → corrective rehabilitation".

[0129] After deforming the initial orthotic insole model, the method also includes: After deforming the initial orthotic insole model, the upper surface of the deformed insole is triangulated using Delaunay 2.5D, and the bottom surface of the orthotic insole is generated by offsetting downwards by a second preset distance along the normal direction of the reference plane; for example, the second preset distance is 1-3mm.

[0130] The deformed upper surface of the insole and the flat bottom surface of the orthopedic insole are combined to generate the target orthopedic insole model.

[0131] In practice, the deformed upper surface of the insole is triangulated using Delaunay 2.5D to generate a mesh model. Then, the reference plane is offset downwards by 3mm along the normal direction to generate the bottom surface of the insole. Next, the boundary vertices of the triangular mesh on the upper surface are connected to the corresponding boundary vertices of the bottom surface to form the side surface. The upper surface, bottom surface, and side surface are then combined into a closed 3D mesh model with solid thickness, for example... Figure 14 As shown: Outer side height: 16.2mm; Inner side height: 21.6mm; Arch height: 13.8mm; Heel thickness: 12.8mm; Reference values ​​for the bottom surface generation range are 3mm, 2mm and 1mm respectively.

[0132] Finally, the model is exported as an STL format file, which can be directly used for 3D printing or CNC machining.

[0133] The orthotic insole model generation system provided by the present invention will be described below. The orthotic insole model generation system described below can be referred to in correspondence with the orthotic insole model generation method described above.

[0134] like Figure 15 As shown, embodiments of the present invention also provide a system for generating orthotic insole models, including: a wearable pressure detection system, an optical scanning device, and a controller; the controller integrates a dual-task machine learning model; Wearable pressure detection systems are used to determine the spatiotemporal feature vector of a user's foot. Optical scanning equipment is used to scan a user's feet to construct a three-dimensional point cloud model of the user's feet; The controller is used to determine the static parameters of the foot and the initial orthotic insole model based on the three-dimensional point cloud model of the user's foot, and to concatenate the static parameters of the foot and the spatiotemporal feature vector of the sole to obtain a multidimensional feature vector; A dual-task machine learning model is used to process the input multidimensional feature vector and output the deformation parameters of the orthopedic insole. The controller is also used to deform the initial orthotic insole model based on the orthotic insole deformation parameters to generate the target orthotic insole model.

[0135] The wearable pressure detection system integrates a distributed flexible sensor array.

[0136] Although the invention has been described herein in conjunction with various embodiments, those skilled in the art will understand and implement other variations of the disclosed embodiments by reviewing the accompanying drawings, the disclosure, and the appended claims in carrying out the claimed invention. In the claims, the word "comprising" does not exclude other components or steps, and "a" or "an" does not exclude a plurality. A single processor or other unit can implement several functions listed in the claims. While different dependent claims may recite certain measures, this does not mean that these measures cannot be combined to produce good results.

[0137] Although the invention has been described in conjunction with specific features and embodiments, it is apparent that various modifications and combinations can be made without departing from the spirit and scope of the invention. Accordingly, this specification and drawings are merely exemplary descriptions of the invention as defined by the appended claims and are to be considered as covering any and all modifications, variations, combinations, or equivalents within the scope of the invention. Clearly, those skilled in the art can make various alterations and modifications to the invention without departing from its spirit and scope. Thus, if such modifications and modifications of the invention fall within the scope of the claims and their equivalents, the invention is also intended to include such modifications and modifications.

Claims

1. A method for generating an orthotic insole model, characterized in that, include: Obtain the user's spatiotemporal feature vector of the foot, static parameters of the foot, and initial orthotic insole model; The multidimensional feature vector is input into a dual-task machine learning model, and the orthopedic insole deformation parameters are output; the multidimensional feature vector is obtained by concatenating the foot static parameters and the foot spatiotemporal feature vector. Based on the deformation parameters of the orthopedic insole, the initial orthopedic insole model is processed to generate the target orthopedic insole model.

2. The method for generating an orthopedic insole model according to claim 1, characterized in that, Before acquiring the user's plantar spatiotemporal feature vector and the three-dimensional point cloud model of the foot, the method further includes: determining the user's plantar spatiotemporal feature vector; Determine the user's plantar spatiotemporal feature vector, including: A wearable pressure detection system is used to collect dynamic pressure time-series data at a preset number of points on the soles of the user's feet; Extract the pressure time integral and pressure center trajectory spatiotemporal features of each point from the preset number of points from the dynamic pressure time series data; the pressure center trajectory spatiotemporal features include at least the envelope area, trajectory length, and average speed of the pressure center point moving within the corresponding time. The pressure time integral and the spatiotemporal features of the pressure center trajectory are determined as the plantar spatiotemporal feature vector.

3. The method for generating an orthopedic insole model according to claim 1, characterized in that, Before acquiring the user's spatiotemporal feature vector of the foot and the 3D point cloud model of the foot, the method further includes: A three-dimensional point cloud model of the foot is constructed, and the static parameters of the foot and the initial orthotic insole model are determined based on the three-dimensional point cloud model of the foot. Constructing a 3D point cloud model of the foot, including: Full-foot multi-view point cloud data were acquired using optical scanning equipment; Multiple feature points are selected from the full-foot multi-view point cloud data, and a high-dimensional descriptor is generated for each of the multiple feature points to obtain multiple high-dimensional descriptors; the multiple feature points include at least points corresponding to high curvature regions, sharp edges, and corners; For each pair of high-dimensional descriptors among the multiple high-dimensional descriptors, the formula is used: ； The Euclidean distance between each pair of high-dimensional descriptors is calculated. Substituting the Euclidean distance into the formula: ； The set of reliable matching point pairs corresponding to the plurality of high-dimensional descriptors is calculated; wherein, For point clouds The first in A high-dimensional descriptor vector of feature points; For point clouds The first in A high-dimensional descriptor vector of feature points; The similarity threshold; A set of reliably matched point pairs; Align each point cloud data in the set of reliable matching point pairs to a unified coordinate system to generate the three-dimensional point cloud model.

4. The method for generating an orthopedic insole model according to claim 3, characterized in that, Determining the static parameters of the foot based on the three-dimensional point cloud model of the foot includes: Identify the position coordinates of multiple marker points in the three-dimensional point cloud model; the multiple marker points include at least the tip of the longest toe, the lowest point of the first metatarsal head, the most convex point of the first metatarsal head, the highest point of the arch, the lowest point of the fifth metatarsal head, the most convex point of the fifth metatarsal head, the lowest point of the center of the heel, and the far end of the foot. The foot static parameters are calculated based on the position coordinates of the multiple marker points; the foot static parameters include foot length, foot width, and arch height.

5. The method for generating an orthopedic insole model according to claim 4, characterized in that, The initial orthotic insole model is determined based on the aforementioned three-dimensional point cloud model of the foot, including: Using the lowest point of the heel center, the lowest point of the first metatarsal head, and the lowest point of the fifth metatarsal head in the three-dimensional point cloud model of the foot as references, a reference plane is obtained by fitting. A cutting plane is generated by offsetting upward by a first preset distance along the normal direction of the reference plane; The foot three-dimensional point cloud model is cut using the cutting plane and all point cloud data below the cutting plane are retained to generate the upper surface curved surface of the orthopedic insole model; The upper surface of the orthopedic insole model is defined as the initial orthopedic insole model.

6. The method for generating an orthopedic insole model according to claim 5, characterized in that, The foot static parameters and the foot spatiotemporal feature vector are input into a dual-task machine learning model to output orthopedic insole deformation parameters. The multidimensional feature vector is learned by using the fully connected layer in the dual-task machine learning model to obtain the learned features. The learned features are classified and identified using the AdaBoost model in the dual-task machine learning model to obtain foot type categories; the foot type categories include at least normal foot, flat foot, and high arch foot. The learned features are processed using the regression network model in the dual-task machine learning model to output the orthopedic insole deformation parameters; the orthopedic insole deformation parameters include the deformation height parameter and the deformation influence range parameter of the area to be deformed.

7. The method for generating an orthopedic insole model according to claim 6, characterized in that, The initial orthotic insole model is processed based on the orthotic insole deformation parameters, including: For the region requiring deformation, a Gaussian decay function is used: ； The upper surface of the initial orthotic insole model is subjected to gradient deformation; wherein, This represents the gradient deformation height at the current point. The deformation height parameter predicted by the regression network model; The Euclidean distance from the deformation center point to the current point; The parameter represents the range of deformation influence.

8. The method for generating an orthopedic insole model according to claim 7, characterized in that, Processing the initial orthotic insole model based on the orthotic insole deformation parameters further includes: After deforming the initial orthotic insole model, the upper surface of the deformed insole is triangulated using Delaunay 2.5D, and the bottom surface of the orthotic insole is generated by offsetting downwards by a second preset distance along the normal direction of the reference plane. The deformed upper surface of the insole and the flat bottom surface of the orthopedic insole are combined to generate the target orthopedic insole model.

9. A system for generating orthotic insole models, characterized in that, include: A wearable pressure detection system, an optical scanning device, and a controller; both the wearable pressure detection system and the optical scanning device are connected to the controller; the controller integrates a dual-task machine learning model; The wearable pressure detection system is used to determine the spatiotemporal feature vector of the user's foot. The optical scanning device is used to scan the user's feet to construct a three-dimensional point cloud model of the user's feet; The controller is used to determine the static parameters of the foot and the initial orthotic insole model based on the three-dimensional point cloud model of the foot, and to concatenate the static parameters of the foot and the spatiotemporal feature vector of the foot to obtain a multidimensional feature vector; The dual-task machine learning model is used to process the input multidimensional feature vector and output the orthopedic insole deformation parameters. The controller is also used to process the initial orthopedic insole model based on the orthopedic insole deformation parameters to generate a target orthopedic insole model.

10. The orthopedic insole model generation system according to claim 9, characterized in that, The wearable pressure detection system integrates a distributed flexible sensor array.

Images