An AI vision-based three-dimensional fabric center positioning method for a sewing machine and a sewing machine

Abstract

The application provides a sewing machine three-dimensional fabric center positioning method and a sewing machine based on AI vision. The method comprises the following steps: acquiring environmental interference data and fabric data in real time and processing the data to obtain vision data; generating point cloud data of the fabric surface; generating a vision positioning result based on the vision data, constructing a three-dimensional model in combination with the point cloud data, generating a model positioning result based on the three-dimensional model, and fusing the vision positioning result and the model positioning result to obtain a current center positioning result of the three-dimensional fabric; and guiding the sewing machine to perform a sewing operation on the three-dimensional fabric based on the current center positioning result. The application generates a vision positioning result and a model positioning result and fuses the results to obtain a current center positioning result, comprehensively combines the intuitive capturing capability of the vision data on the spatial features and the dynamic adaptation advantage of the three-dimensional model on the physical form, overcomes the positioning deviation caused by the isolated processing of the vision and physical data in the prior art, and significantly improves the accuracy of the three-dimensional fabric center positioning.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of intelligent sewing, in particular, to an AI vision-based center positioning method for three-dimensional fabric of a sewing machine and the sewing machine. BACKGROUND

[0002] In the sewing process of three-dimensional fabric (such as collar, cuff, shoes, hats and other clothing with three-dimensional curved surface), the center positioning accuracy is a key factor determining the sewing quality. However, in the existing technology, there are problems of poor positioning accuracy and large sewing deviation in the center positioning and sewing process of three-dimensional fabric. Three-dimensional fabric has flexible characteristics, and its shape is easily deformed by external forces such as pulling and pressing during the sewing process. However, the existing positioning methods are mostly based on static images or simple physical sensing data, which are difficult to capture and adapt to such dynamic deformation in real time. For example, traditional visual positioning only extracts features from a single frame or limited frames of images, which cannot reflect the real-time shape changes of the fabric due to stress, resulting in a lag of the positioning result from the actual center position. Although physical sensing positioning can detect the stress, it lacks accurate description of the spatial shape of the fabric, making it difficult to determine the center coordinates after deformation, and ultimately causing positioning deviation. SUMMARY

[0003] Based on the problems existing in the prior art, the present application provides an AI vision-based center positioning method for three-dimensional fabric of a sewing machine and the sewing machine. The specific scheme is as follows:

[0004] Firstly, the present application provides an AI vision-based center positioning method for three-dimensional fabric of a sewing machine, which comprises:

[0005] Real-time acquisition of environmental interference data, image data and physical state data of three-dimensional fabric, and anti-interference optimization of the image data based on the environmental interference data to obtain visual data;

[0006] Extracting feature data of three-dimensional fabric from the visual data by using a preset AI recognition algorithm, generating three-dimensional point cloud of the fabric surface after three-dimensional reconstruction based on the feature data, and dynamically correcting the three-dimensional point cloud based on the physical state data to obtain point cloud data;

[0007] Generating a visual positioning result based on the visual data, constructing a three-dimensional model dynamically adapted to the real-time physical shape of the fabric in combination with the point cloud data, and generating a model positioning result based on the three-dimensional model, and fusing the visual positioning result and the model positioning result to obtain a current center positioning result of the three-dimensional fabric;

[0008] Guiding the sewing machine to perform the sewing operation of the three-dimensional fabric based on the current center positioning result.

[0009] In some embodiments, further comprising: predicting the deformation direction and amplitude of the cloth in a subsequent period by analyzing the timing change rule of the physical state data, and further inferring the pre-judgment center positioning result of the three-dimensional cloth;

[0010] Guiding the sewing machine to perform the sewing operation of the three-dimensional cloth based on the current center positioning result and the pre-judgment center positioning result.

[0011] In some embodiments, further comprising: calculating the confidence of the current center positioning result and the pre-judgment center positioning result, respectively;

[0012] According to the real-time sewing speed of the sewing machine and the deformation rate of the three-dimensional cloth, dynamically allocating the fusion weight of the current center positioning result and the pre-judgment center positioning result; wherein the fusion weight is positively correlated with the confidence of the corresponding positioning result;

[0013] Based on the fusion weight, the current center positioning result and the pre-judgment center positioning result are weighted and fused to generate a final center positioning result to guide the sewing machine to perform the sewing operation.

[0014] In some embodiments, the environmental interference data includes environmental light intensity, vibration frequency data and vibration amplitude data; the image data includes a pair of binocular images synchronously collected;

[0015] The anti-interference optimization specifically includes: based on the environmental light intensity, performing adaptive optical correction on the pair of binocular images; based on the vibration frequency data and the vibration amplitude data, performing geometric offset compensation on the optically corrected images, and realizing coordinate correction by calculating the offset compensation value of each feature point in the image.

[0016] In some embodiments, the pre-judgment center positioning result is obtained by:

[0017] Obtaining physical state data in a continuous period, analyzing the change trend and fluctuation range of the physical state data in the continuous period, and determining the dynamic deformation characteristics of the cloth;

[0018] Inputting the historical timing data composed of the dynamic deformation characteristics and the physical state data into a preset timing prediction model to predict the deformation direction and amplitude of the three-dimensional cloth in a subsequent period;

[0019] Based on the current center positioning result of the three-dimensional cloth, combining the output deformation direction and amplitude, the pre-judgment center positioning result of the three-dimensional cloth in the subsequent period is calculated.

[0020] In some embodiments, the visual positioning result is generated based on the visual data, comprising:

[0021] Extract edge feature points and texture feature points of the three-dimensional fabric from the visual data, filter out effective feature points by judging whether the spatial distance between the feature points and the surrounding feature points is within a preset range, calculate the geometric center of all effective feature points based on the three-dimensional coordinates of the effective feature points, and take the geometric center as the visual positioning result.

[0022] In some embodiments, the model positioning result is generated based on the three-dimensional model, including:

[0023] Extract the symmetry surface and top surface features of the model from the three-dimensional model, calculate the intersection line of the symmetry surface and the geometric center of the top surface, and combine the material density distribution characteristics of the three-dimensional model to weight and fuse the intersection line and the geometric center of the top surface; take the fused coordinates as the model positioning result; wherein the features corresponding to the areas with higher material density have higher weights in the weighted fusion than the features corresponding to the areas with lower material density.

[0024] In some embodiments, the construction process of the three-dimensional model includes:

[0025] Preprocess the point cloud data to retain effective point clouds representing the surface morphology of the fabric;

[0026] Re-establish an initial three-dimensional model of the three-dimensional fabric based on the effective point clouds;

[0027] Combine the real-time physical state data to dynamically adjust the initial three-dimensional model; wherein when the physical state data shows that the fabric has stretched, adjust the size of the corresponding area of the model in the stretching direction; when the physical state data shows that the fabric has wrinkles, generate a convex or concave structure matching the wrinkle morphology in the corresponding area of the model;

[0028] Every interval of a preset period, re-adjust the three-dimensional model according to the updated point cloud data and physical state data, so that the morphology of the three-dimensional model always matches the real-time physical morphology of the fabric, achieving dynamic adaptation.

[0029] Secondly, the application proposes an AI visual sewing machine three-dimensional fabric center positioning system for realizing the AI visual sewing machine three-dimensional fabric center positioning method of the first part, including:

[0030] An input unit is configured to acquire environmental interference data, image data and physical state data of the three-dimensional fabric in real time, and to obtain visual data by anti-interference optimization of the image data based on the environmental interference data.

[0031] a three-dimensional reconstruction unit configured to extract feature data of the three-dimensional fabric from the visual data by using a preset AI recognition algorithm, generate a three-dimensional point cloud of a fabric surface based on three-dimensional reconstruction of the feature data, and dynamically correct the three-dimensional point cloud based on the physical state data to obtain point cloud data;

[0032] a center positioning unit configured to generate a visual positioning result based on the visual data, construct a three-dimensional model dynamically adapted to a real-time physical form of the fabric in combination with the point cloud data, generate a model positioning result based on the three-dimensional model, and fuse the visual positioning result and the model positioning result to obtain a current center positioning result of the three-dimensional fabric;

[0033] an output unit configured to guide the sewing machine to perform a sewing operation of the three-dimensional fabric based on the current center positioning result.

[0034] In a third part, the application provides an intelligent sewing machine for implementing the AI visual sewing machine three-dimensional fabric center positioning method of any one of the first part.

[0035] Beneficial effects: The application provides an AI visual sewing machine three-dimensional fabric center positioning method and a sewing machine. The current center positioning result is obtained by generating a visual positioning result and a model positioning result and fusing them. The method combines the intuitive spatial feature capturing capability of visual data and the dynamic adaptation advantage of a three-dimensional model to the physical form, overcomes the positioning deviation caused by isolated processing of visual data and physical data in the prior art, significantly improves the accuracy of three-dimensional fabric center positioning, and enables the sewing operation to match the dynamic form change of the three-dimensional fabric in real time, effectively reduces the deviation of the sewing stitch from the preset center, improves the sewing precision of the three-dimensional fabric, and meets the demand for positioning accuracy in the precision sewing scene.

[0036] To make the above objectives, characteristics and advantages of the application more obvious and easy to understand, the following preferred embodiments are described in detail below, and the accompanying drawings are referred to as follows. BRIEF DESCRIPTION OF DRAWINGS

[0037] To more clearly illustrate the technical solutions of the embodiments of the application, the following will briefly introduce the drawings needed to be used in the embodiments. It should be understood that the following drawings only show some of the embodiments of the application, and therefore should not be considered as a limitation to the scope. For those skilled in the art, other related drawings can also be obtained without creative labor based on these drawings.

[0038] Figure 1 is a flowchart of the sewing machine three-dimensional fabric center positioning method of the application;

[0039] Figure 2 is a principle diagram of the sewing machine three-dimensional fabric center positioning method of the application;

[0040] Figure 3 is the final center positioning result acquisition process schematic diagram of the present application;

[0041] Figure 4 is the model positioning result acquisition process schematic diagram of the present application;

[0042] Figure 5 is the three-dimensional fabric center positioning system module schematic diagram of the sewing machine of the present application.

[0043] Reference signs: 1-input unit; 2-three-dimensional reconstruction unit; 3-center positioning unit; 4-output unit. DETAILED DESCRIPTION

[0044] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0045] The present application proposes an AI vision-based three-dimensional fabric center positioning method for a sewing machine. By integrating advanced recognition technology and intelligent control algorithms, the production efficiency and product quality are improved, and the production cost is reduced, which has significant economic and social benefits. The flowchart of the AI vision-based three-dimensional fabric center positioning method for the sewing machine is shown in the accompanying Figure 1 The specific scheme is as follows:

[0046] An AI vision-based three-dimensional fabric center positioning method for a sewing machine, comprising:

[0047] 101, real-time acquisition of environmental interference data, image data and physical state data of the three-dimensional fabric, anti-interference optimization of the image data based on the environmental interference data to obtain visual data;

[0048] 102, extracting feature data of the three-dimensional fabric from the visual data using a preset AI recognition algorithm, generating a three-dimensional point cloud of the fabric surface after three-dimensional reconstruction based on the feature data, and dynamically correcting the three-dimensional point cloud based on the physical state data to obtain point cloud data;

[0049] 103, generating a visual positioning result based on the visual data, constructing a three-dimensional model dynamically adapted to the real-time physical form of the fabric in combination with the point cloud data, generating a model positioning result based on the three-dimensional model, and fusing the visual positioning result and the model positioning result to obtain a current center positioning result of the three-dimensional fabric;

[0050] 104, guiding the sewing machine to perform a sewing operation on the three-dimensional fabric based on the current center positioning result.

[0051] In some embodiments, the sewing operation includes:

[0052] Firstly, the AI vision system synchronously collects image data of the cloth through binocular cameras, obtains environmental light intensity and vibration data through environmental sensors, and collects stress degree and deformation degree of the cloth through physical sensors. Based on the environmental interference data, the image data is optimized, the edge feature points, texture feature points and surface relief features of the cloth are extracted, the three-dimensional coordinates are calculated after screening the effective feature points, and the spatial profile, center position and wrinkle convexity / stacking depression area of the cloth are determined. The real-time shape model of the cloth is constructed through the three-dimensional coordinates of the feature points, and the key parameters such as the cloth center coordinates, cloth surface flatness, elastic / plastic deformation area distribution, wrinkle depth and stacking height are determined.

[0053] Subsequently, the AI vision system converts the analyzed visual data and physical data into instruction parameters recognizable by the robot, including spatial positioning instructions, posture adjustment instructions and force control instructions. The spatial positioning instruction clearly indicates the target coordinates of the robot's grasp, which is based on the cloth center positioning result or the preset three-dimensional coordinates of the grasp point, including the accurate positions of x / y / z axes, to ensure that the robot accurately grasps the core area. The posture adjustment instruction is to adapt the robot's posture to the cloth shape, such as using a horizontal grasping posture when the cloth is flat, adjusting the robot's angle when there are wrinkles to avoid pulling. The force control instruction is to set the clamping force of the robot's gripper based on the cloth material information and deformation degree. The elastic deformation area uses light clamping to avoid excessive stretching, and the plastic deformation area uses flexible lifting + point touch clamping to avoid damaging the fixed shape.

[0054] Step 101 is the basic data processing link of the entire positioning method. By synchronously collecting multiple types of data that affect positioning, and specifically eliminating the influence of environmental interference on visual information, while obtaining key data reflecting the physical state of the cloth. This step can avoid image distortion caused by environmental factors, ensure the integrity and accuracy of the cloth features in the visual data, and at the same time, the synchronous acquisition of physical state data provides a data basis for subsequent processing of cloth dynamic deformation, solving the problem of unreliable basic data caused by environmental interference and lack of physical data in traditional positioning.

[0055] The environmental interference data refers to environmental factor data affecting image acquisition quality, such as illumination intensity (representing environmental light and dark changes) and device vibration parameters (representing vibration amplitude and frequency when the sewing machine is running), which can cause image blurring, deviation or brightness abnormalities, and directly affect the accuracy of subsequent feature extraction. The image data is the visual information of the three-dimensional fabric obtained by a visual acquisition device (such as a binocular camera), which includes spatial features such as the edges, textures and surface morphology of the fabric, and provides the original visual basis for three-dimensional positioning. The physical state data is quantitative data reflecting the stress and deformation of the fabric, such as the tension and pressure (representing the strength of external force) and the degree of deformation (representing the size of the fabric morphology change). Since the three-dimensional fabric is flexible, it is prone to stretching, wrinkling and other deformations under stress, and these data are the key basis for subsequent deformation adaptation.

[0056] The anti-interference optimization of the image data based on the environmental interference data essentially adjusts the image parameters to offset the influence of light and corrects the image coordinates to offset the deviation caused by vibration, ultimately obtaining visual data that can clearly reflect the features of the fabric. This step avoids image distortion caused by environmental factors and provides high-quality input for subsequent feature extraction. Meanwhile, the acquisition of physical state data provides a basis for adapting to the dynamic deformation of the fabric.

[0057] Step 102 is the core step of constructing an accurate three-dimensional representation. First, the key features of the fabric are extracted from the visual data, then a three-dimensional point cloud is reconstructed based on these features, and finally the point cloud is corrected in combination with the physical state data to adapt to the actual physical deformation of the fabric. By generating point cloud data that can truly reflect the real-time physical morphology of the fabric, the problem of traditional three-dimensional point cloud relying solely on vision and being unable to adapt to the deformation of flexible fabric is solved, providing a reliable three-dimensional representation basis for subsequent construction of an accurate three-dimensional model.

[0058] A predetermined AI recognition algorithm is used to extract feature data of the three-dimensional fabric from the visual data, such as edges, textures, surface undulations and other feature information that can represent the fabric morphology. Based on these features, a three-dimensional point cloud of the fabric surface is generated through three-dimensional reconstruction techniques such as stereo matching and depth calculation, which preliminarily reflects the spatial morphology of the fabric. Since the fabric is a flexible material, it will deform under stress, so the three-dimensional point cloud needs to be dynamically corrected based on the physical state data to make the point cloud fit the actual physical morphology of the fabric. This step solves the problem of point cloud generated based solely on vision being unable to reflect the deformation of the fabric under stress, and lays the foundation for subsequent construction of an accurate three-dimensional model.

[0059] The preset AI recognition algorithm is an artificial intelligence algorithm for extracting fabric features from visual data. Its role is to accurately identify and extract key information that can represent the fabric form by learning the visual feature patterns of the fabric, overcoming the poor adaptability of traditional feature extraction methods to complex textures and flexible deformation. The feature data is the fabric features extracted by the AI algorithm, including edge contour (defining the fabric boundary), surface texture (reflecting fabric material details), and surface relief (representing three-dimensional forms such as wrinkles and protrusions). These data are the basis for three-dimensional reconstruction. The three-dimensional point cloud is a set of three-dimensional space points generated based on the spatial coordinates of the feature data. Each point represents a position on the fabric surface, and the whole constitutes the three-dimensional form of the fabric surface. Dynamic correction is the adjustment process of the three-dimensional point cloud based on physical state data. Since the initial three-dimensional point cloud only reflects the visual form, the actual form may deviate from the visual form due to the deformation of the fabric under stress. By adjusting the coordinates of the corresponding points in the point cloud based on physical state data, the point cloud is fitted to the actual physical form of the fabric, and the point cloud data is obtained.

[0060] Step 103 is the core step to achieve center positioning. By fusing two different dimensional positioning results, the intuitive visual information and the adaptability of the three-dimensional model to the physical form are integrated to obtain an accurate current center positioning result, which can significantly improve the accuracy of three-dimensional fabric center positioning, especially when the fabric undergoes dynamic deformation, the accuracy of positioning can still be maintained, solving the problem of poor accuracy of traditional single positioning method in deformation scenarios.

[0061] Among them, the visual positioning result is the center position of the fabric based on visual data. The principle is to extract visual features that can reflect the center from visual data, including geometric center, symmetric center, etc., by calculating the spatial coordinates of these features. Its advantage is that it is directly based on visual information and can quickly reflect the apparent center of the fabric. The three-dimensional model is a digital model that can dynamically adapt to the real-time physical form of the fabric based on point cloud data. It not only contains the three-dimensional contour of the fabric, but also can update synchronously with the deformation of the fabric such as stretching and wrinkling, ensuring that the model form is consistent with the actual form of the fabric.

[0062] The model positioning result is the center position of the fabric based on the three-dimensional model. The principle is to extract features that can reflect the physical center from the three-dimensional model. Its advantage is that it combines the physical form of the fabric and is more stable in deformation scenarios. Since the two positioning results have their own advantages, visual positioning is intuitive and model positioning is adaptive to deformation. By fusing them, the limitations of single method can be offset, and a more accurate current center positioning result can be obtained.

[0063] Specifically, the visual features reflecting the center of the cloth are extracted from the visual data, and the preliminary center position is calculated. At the same time, a three-dimensional model is constructed combined with the corrected point cloud data, which can dynamically adapt to the real-time physical form of the cloth, such as synchronous changes with the stretching and wrinkling of the cloth. Based on the three-dimensional model, features reflecting the physical center of the cloth are extracted to generate the model positioning result. Finally, the visual positioning result and the model positioning result are fused to overcome the limitations of a single method, and the current center positioning result of the three-dimensional cloth is obtained.

[0064] Step 104 is the step of converting the positioning result into actual sewing operation. Based on the current center positioning result, the motion trajectory of the sewing machine head is controlled to keep alignment with the center of the cloth, ensuring accurate sewing position. The accurate positioning result is converted into high-quality sewing effect, avoiding problems such as sewing misalignment and needle trace skew caused by positioning deviation, directly improving the sewing quality of the three-dimensional cloth and meeting the needs of high-precision sewing scenarios. In actual application, the current center positioning result is converted into control parameters of the sewing machine, such as displacement and speed of the machine head, and the sewing machine is driven by the control system to perform sewing operation, making the needle trace follow the center of the cloth or a preset path. The effect of this step is to convert the accurate positioning result into high-quality sewing effect, avoiding problems such as sewing misalignment and needle trace skew caused by positioning deviation, and improving the sewing quality of the three-dimensional cloth.

[0065] To further adapt to the dynamic deformation of the three-dimensional cloth during the sewing process and compensate for the response delay of the sewing machine, in some specific embodiments, it also includes: predicting the deformation direction and amplitude of the cloth in the subsequent period by analyzing the time sequence change law of the physical state data, and then predicting the center positioning result of the three-dimensional cloth; based on the current center positioning result and the predicted center positioning result, guiding the sewing machine to perform the sewing operation of the three-dimensional cloth. The time sequence change law is the change trend of the physical state data collected at consecutive time points over time. By analyzing these time sequence laws, the dynamic deformation features of the cloth can be extracted, which can reflect the inherent laws and trends of cloth deformation. The preset time sequence prediction model is a time sequence prediction algorithm trained based on a large number of historical deformation samples, including LSTM, time sequence Transformer, etc., which is based on historical time sequence physical state data and dynamic deformation features to learn the evolution law of cloth deformation. After inputting the extracted dynamic deformation features and historical physical state data into the model, the model can output the deformation direction and amplitude of the cloth in the subsequent preset period.

[0066] The core role of the pre-judgment center positioning result is to solve the contradiction between the dynamic deformation of the cloth and the response delay of the equipment in the sewing process, to provide a forward-looking control reference for the sewing machine, so as to maintain the sewing precision under high speed or strong deformation scene. The three-dimensional cloth will continuously deform due to external forces such as pulling and pressing during sewing, and there is a physical response delay in adjusting the position of the sewing machine head. If only relying on the current center positioning result, when the cloth deforms significantly within the delay time, the position reached by the machine head will deviate from the actual center of the cloth. The pre-judgment center positioning result predicts the deformation direction and amplitude of the future period in advance by analyzing the time sequence law of the physical state data, which is essentially a reasonable deduction of the future state based on the historical deformation law, which is equivalent to providing an advance for the sewing machine, so that it can pre-adjust the position of the machine head to offset the error caused by the response delay.

[0067] The current center positioning result reflects the actual center position of the cloth at the current time, which is calculated based on real-time collected visual data and physical state data. The pre-judgment center positioning result is the estimated center position after a future preset period, which is derived based on historical time series data and prediction model. The current center positioning result is a real-time reference, which ensures accurate capture of the current state; the pre-judgment center positioning result is a forward-looking compensation, which is used to offset the response delay of the equipment, and the two work together to achieve high-precision control in dynamic scenarios.

[0068] The acquisition process of the pre-judgment center positioning result combines time sequence analysis and prediction model to realize the forward-looking capture of the dynamic deformation of the three-dimensional cloth. Further, the acquisition process of the pre-judgment center positioning result specifically includes: acquiring physical state data in a continuous period, analyzing the change trend and fluctuation range of the physical state data in the continuous period, and determining the dynamic deformation characteristics of the cloth; inputting the historical time series data composed of the dynamic deformation characteristics and the physical state data into a preset time sequence prediction model to predict the deformation direction and amplitude of the three-dimensional cloth in the subsequent period; based on the current center positioning result of the three-dimensional cloth, combining the output deformation direction and amplitude, the pre-judgment center positioning result of the three-dimensional cloth in the subsequent period is calculated. By mining the time sequence law of the physical state data, the influence of cloth deformation on the center position is predicted in advance, solving the problem that the traditional positioning only relies on the current state and cannot cope with the positioning lagging behind the actual deformation in high-speed sewing or strong deformation scenarios, so that the sewing machine can adjust the position of the machine head in advance, further improving the sewing precision in dynamic scenarios.

[0069] First, physical state data in a continuous time period is obtained. The continuous time period is usually a time window matching the cloth deformation response speed, ensuring that the evolution process of a deformation can be captured completely; the physical state data includes quantitative data such as cloth stress values and deformation rates collected continuously in the time period. By analyzing the change trend and fluctuation range of the physical state data in the continuous time period, dynamic deformation characteristics of the cloth can be extracted. These characteristics include the dominant type of deformation, the main direction of deformation, the change rate and stability, providing a feature basis for subsequent prediction.

[0070] Next, the historical time series data composed of the dynamic deformation characteristics and the physical state data is input into a preset time series prediction model. The historical time series data is a collection of physical state data and corresponding dynamic deformation characteristics arranged in chronological order, containing the historical evolution law of cloth deformation. The model can output the deformation direction and amplitude of the three-dimensional cloth in a subsequent preset time period by analyzing the change law of the dynamic deformation characteristics in the historical time series data and combining the numerical change of the physical state data, realizing quantitative prediction of future deformation.

[0071] Finally, based on the current center positioning result of the three-dimensional cloth, the deformation direction and amplitude output are combined to calculate the subsequent time period prediction center positioning result. Specifically, taking the three-dimensional coordinates of the current center positioning result as the reference, the coordinates are offset according to the predicted deformation direction and amplitude, thereby obtaining the estimated position of the cloth center in the subsequent time period.

[0072] To further improve the robustness of center positioning in dynamic scenarios, in some specific embodiments, it also includes: calculating the confidence of the current center positioning result and the prediction center positioning result respectively; dynamically assigning the fusion weight of the current center positioning result and the prediction center positioning result according to the real-time sewing speed of the sewing machine and the deformation rate of the three-dimensional cloth; wherein the fusion weight is positively correlated with the confidence of the corresponding positioning result; based on the fusion weight, the current center positioning result and the prediction center positioning result are weighted and fused to generate the final center positioning result to guide the sewing machine to perform the sewing operation. By dynamically balancing the real-time accuracy of the current positioning and the forward compensation of the prediction positioning, the problem of insufficient accuracy of traditional fixed weight fusion in high-speed sewing or strong deformation scenarios is solved, which can rely on the current result with high confidence to ensure real-time accuracy at low speed and low deformation, and can offset the device delay through the prediction result with high confidence at high speed and high deformation, finally realizing high accuracy and high stability of three-dimensional cloth center positioning in all scenarios. The acquisition process of the final center positioning result is shown in FIG. 8. Figure 3

[0073] ​First, the confidence of the current center positioning result and the prediction center positioning result is calculated respectively. The confidence is a quantitative index for measuring the reliability of the positioning result: the confidence of the current center positioning result is mainly determined based on the quality of real-time data, for example, the number of effective feature points in the visual data, the more the number of feature extraction is more reliable; the matching degree of point cloud data and physical state data, the smaller the deviation, the more accurate the modeling, the higher the data quality, the higher the confidence of the current positioning result. The confidence of the prediction center positioning result is related to the regularity of historical time series data, for example, whether the physical state data changes smoothly; it is also related to the historical prediction error of the preset time series prediction model, the smaller the error, the stronger the model generalization ability, the more stable the historical regularity, the smaller the model error, and the higher the confidence of the prediction result.

[0074] Then, according to the real-time sewing speed of the sewing machine and the deformation rate of the three-dimensional fabric, the fusion weight of the two is dynamically allocated, and the fusion weight is positively correlated with the confidence of the positioning result corresponding thereto. The real-time sewing speed reflects the urgency of the device response, the faster the speed, the shorter the time window of the head adjustment, the greater the influence of the response delay, at this time, if the prediction result confidence is high, its weight needs to be increased to compensate in advance. The deformation rate of the three-dimensional fabric reflects the degree of change in the fabric form, the higher the deformation rate, the shorter the timeliness of the current positioning result, which will soon be invalid due to deformation, if the prediction can effectively capture the deformation trend, its weight should be increased, otherwise it needs to be reduced to avoid the interference of false prediction. At the same time, the positive correlation between the weight and the confidence ensures that the positioning result with higher reliability dominates in the fusion, for example, when the confidence of the current result is 0.8 and the confidence of the prediction result is 0.6, the weight of the former is higher than that of the latter, avoiding the interference of the low confidence result on the final positioning.

[0075] Finally, the current center positioning result and the prediction center positioning result are weighted and fused based on the fusion weight to generate the final center positioning result. Specifically, the three-dimensional coordinates of the current positioning result and the three-dimensional coordinates of the prediction positioning result are multiplied by their respective weights and summed, such as the final coordinates equal to the current coordinates multiplied by the current weight plus the prediction coordinates multiplied by the prediction weight, the final positioning reference considering real-time and foresight is obtained to guide the sewing machine to perform the sewing operation.

[0076] In some embodiments, the environmental interference data includes ambient light intensity, vibration frequency data, and vibration amplitude data; the image data includes a pair of binocular images synchronously collected; and the anti-interference optimization specifically includes: adaptive optical correction of the pair of binocular images based on the ambient light intensity; and geometric offset compensation of the optically corrected images based on the vibration frequency data and the vibration amplitude data, with the offset compensation value of each feature point in the image being calculated to correct the coordinates. The ambient light intensity reflects the brightness of the shooting environment, and too strong light will cause overexposure of the image highlights and loss of fabric texture, and too weak light will cause the image dark features to be blurred. The vibration frequency data and the vibration amplitude data represent the mechanical vibration characteristics of the sewing machine during operation, and the vibration will cause periodic or random offset of the images captured by the camera, so that the position of the fabric feature points in the image does not match the actual position. Through the anti-interference optimization, the interference of light and vibration on the image data can be effectively weakened, and the completeness and position accuracy of the fabric features in the obtained visual data are significantly improved, laying a reliable foundation for the subsequent feature extraction, three-dimensional reconstruction, and positioning result accuracy.

[0077] The image data uses a pair of binocular images synchronously collected, which are obtained by simultaneously shooting a stereoscopic fabric from different angles by two cameras with consistent parameters. The three-dimensional coordinates of the fabric surface feature points can be calculated through the disparity between the left and right images, providing original depth information for subsequent three-dimensional reconstruction. Synchronous collection can avoid the problem of mismatching of dynamic fabric morphology caused by time difference.

[0078] The anti-interference optimization specifically includes two parts: one is adaptive optical correction of the pair of binocular images based on the ambient light intensity. When the ambient light intensity is low, the image brightness is increased, the contrast is increased, and the noise is suppressed to enhance the distinguishability of the fabric edges and texture. When the light intensity is too high, the image exposure is reduced and the highlights are suppressed to avoid the loss of features in the light-reflecting area of the fabric surface, so that the key features of the fabric can be clearly presented in the image under different light conditions. The second is geometric offset compensation of the optically corrected images based on the vibration frequency data and the vibration amplitude data. By analyzing the vibration frequency and amplitude, the corresponding relationship between the vibration and the offset of the image feature points is established, the offset compensation value of each feature point in the image due to vibration is calculated, and the coordinates of the feature points are corrected according to the compensation value to eliminate the position deviation of the feature points caused by vibration and ensure that the spatial position of the fabric features in the image is consistent with the actual position.

[0079] In some specific embodiments, the generating of the visual positioning result based on the visual data comprises: extracting edge feature points and texture feature points of the three-dimensional cloth from the visual data, screening out effective feature points by judging whether the spatial distance between the feature points and the surrounding feature points is within a preset range, calculating the geometric center of all the effective feature points based on the three-dimensional coordinates of the effective feature points, and taking the geometric center as the visual positioning result. Relying on the intuitive visual features of the cloth, the spatial center position of the cloth can be quickly and directly reflected, providing a reliable visual reference for subsequent fusion with the model positioning result and improving the overall positioning accuracy.

[0080] The edge feature points refer to key point positions on the profile boundary of the three-dimensional cloth, such as inflection points and arc vertices of the cloth edge, which can clearly define the spatial profile range of the cloth and are the basis for judging the overall shape of the cloth; the texture feature points are point positions corresponding to unique textures formed on the cloth surface due to material and weaving method, which are widely distributed and can assist in correcting possible local deviations of the edge feature points, and the combination of the two can comprehensively cover the spatial features of the cloth.

[0081] The effective feature points are screened out by judging whether the spatial distance between the feature points and the surrounding feature points is within a preset range. Since the visual data may have noise or local occlusion, some of the extracted feature points are not effective information that truly reflects the shape of the cloth. The preset range is set according to the actual size and texture density of the cloth. If the spatial distance between a feature point and multiple surrounding feature points exceeds the range, it means that the point may be a noise or distortion point and needs to be removed; only the feature points with spatial distance within the preset range are determined as effective feature points, thereby ensuring the reliability of the feature points participating in the positioning calculation. The three-dimensional coordinates of the effective feature points are calculated by the parallax principle of binocular vision, containing spatial position information of three dimensions of x, y and z; the calculation method of the geometric center is to take the average of the x coordinates, y coordinates and z coordinates of all effective feature points, and the average coordinates obtained are the three-dimensional coordinates of the geometric center. The geometric center can intuitively reflect the overall center position of the three-dimensional cloth in the visual space, and since it is calculated based on the screened effective feature points, it effectively avoids the interference of noise and distortion points on the positioning result, ensuring the accuracy of the visual positioning result.

[0082] The process of generating the model positioning result based on the three-dimensional model extracts the structural features and material characteristics of the model and fuses them to obtain a positioning result that reflects the physical center of the fabric. In some specific embodiments, generating the model positioning result based on the three-dimensional model includes extracting the symmetrical surface and top surface features of the model from the three-dimensional model, calculating the intersection line of the symmetrical surface and the geometric center of the top surface, and combining the material density distribution characteristics of the three-dimensional model to weight and fuse the intersection line and the geometric center of the top surface. The fused coordinates are used as the model positioning result. In this process, the features corresponding to the areas with higher material density have higher weights in the weighting fusion than the features corresponding to the areas with lower material density. The process is shown in FIG. 8. Figure 4

[0083] First, the symmetrical surface and top surface features of the model are extracted from the three-dimensional model. The symmetrical surface refers to a plane in the three-dimensional model that reflects the symmetry of the fabric structure, such as a longitudinal symmetrical surface for a fabric that is symmetrical left and right, and a horizontal symmetrical surface for a fabric that is symmetrical front and back. These symmetrical surfaces are determined by the spatial distribution of the model surface feature points, and their core function is to reflect the structural symmetry reference of the fabric in three-dimensional space. The top surface feature is the surface feature of the three-dimensional model facing the sewing operation side, which includes the profile, relief, and other morphological information of the surface. As the main action surface for sewing operations, its features can directly relate to the center position of the operation requirements.

[0084] Next, the intersection line of the symmetrical surface and the geometric center of the top surface are calculated. The intersection line of the symmetrical surface is a straight line formed by the intersection of multiple symmetrical surfaces, such as the intersection line of the longitudinal symmetrical surface and the horizontal symmetrical surface, which usually corresponds to the structural center axis of the fabric and is defined as a center reference from the overall structure perspective. The geometric center of the top surface is obtained by calculating the average value of the three-dimensional coordinates of the top surface feature points, reflecting the center position of the top surface in space and being defined as a center reference from the operation surface perspective.

[0085] Then, the intersection line and the geometric center of the top surface are weighted and fused in combination with the material density distribution characteristics of the three-dimensional model. The material density distribution characteristics refer to the tightness of the material in different areas of the three-dimensional model, such as areas with higher density due to tight weaving and fiber arrangement, and areas with lower density due to looseness. Since the areas with higher density have more stable structures and are less affected by deformation, the features corresponding to these areas (such as the symmetrical surface intersection line segments or top surface feature points located in the high-density areas) are more reliable in positioning. Therefore, in the weighting fusion, the weights of these features are higher than those of the features corresponding to the areas with lower density. In the specific fusion process, the weights are assigned according to the material density of the areas where the symmetrical surface intersection line and the geometric center of the top surface are located, with higher weights for higher density areas. The fused coordinates are obtained through weighted calculation.

[0086] ​Finally, the fused coordinates are taken as the model positioning result. This result integrates the structural symmetry of the cloth (reflected by the intersection of the symmetry planes) and the center of the operation surface (reflected by the geometric center of the top surface), and strengthens the influence of the stable area features with the help of the material density characteristics, avoiding the positioning deviation of a single structure or surface feature in the deformation scenario, so that the model positioning result can more accurately reflect the physical center of the three-dimensional cloth, providing a reliable physical form benchmark for subsequent fusion with the visual positioning result.

[0087] The construction process of the three-dimensional model is carried out around accurate modeling-dynamic adaptation, combining point cloud processing with real-time adjustment to ensure that the model is consistent with the real-time physical form of the cloth. In some specific embodiments, the construction process of the three-dimensional model includes: preprocessing the point cloud data to retain effective point clouds representing the surface form of the cloth; based on the effective point clouds, rebuilding an initial three-dimensional model of the three-dimensional cloth; combining real-time physical state data, dynamically adjusting the initial three-dimensional model; wherein when the physical state data shows that the cloth has stretched, the size of the corresponding area of the model is adjusted along the stretching direction; when the physical state data shows that the cloth has wrinkles, a protrusion or depression structure matching the wrinkle form is generated in the corresponding area of the model; every interval of a preset period, the three-dimensional model is adjusted again according to the updated point cloud data and physical state data, so that the form of the three-dimensional model always remains consistent with the real-time physical form of the cloth, realizing dynamic adaptation. Through this process, the three-dimensional model can accurately reflect the three-dimensional form of the cloth and can follow its dynamic deformation in real time, providing a highly matched digital benchmark for the physical form of the cloth for subsequent generation of reliable model positioning results, solving the problem that traditional static models cannot adapt to the dynamic changes of flexible cloth.

[0088] The point cloud data may contain noise points generated by environmental interference, as well as background points that do not belong to the cloth, which will interfere with the accuracy of the model. Preprocessing identifies and removes noise points through filtering algorithms, separates and removes background points through threshold segmentation, and finally retains only effective point clouds that can truly reflect the form features of the cloth surface, such as undulations, edges, and textures, providing pure basic data for subsequent modeling. Based on the effective point clouds, an initial three-dimensional model of the three-dimensional cloth is rebuilt. The initial three-dimensional model is a continuous three-dimensional surface structure converted from discrete effective point clouds through surface reconstruction technology, which can initially present the overall contour, surface curvature, and main form features of the cloth, forming a digital model framework that basically matches the current state of the cloth.

[0089] The physical state data can reflect the dynamic changes of the cloth in real time: when the data shows that the cloth is stretched, the model adjusts the size of the corresponding area in the stretching direction in proportion to the degree of deformation, ensuring that the size of the model is consistent with the actual size of the stretched cloth; when the data shows that the cloth has wrinkles, the model generates a convex or concave structure matching the depth and range of the wrinkles in the corresponding area, and restores the three-dimensional shape of the wrinkles by adjusting the height or curvature of the local grid. This adjustment upgrades the model from a static initial state to a dynamic state that conforms to real-time deformation.

[0090] Every interval, the three-dimensional model is adjusted based on the updated point cloud data and physical state data. The preset interval is usually synchronized with the image and physical data acquisition frequency (such as once every 0.02 seconds), ensuring that the model update speed can keep up with the deformation speed of the cloth. Each time the model is updated, the new point cloud data provides the latest surface morphology details, and the new physical state data provides the latest deformation trend, both of which drive the model to continuously adjust, so that the features such as the contour, size, and wrinkles of the model are always consistent with the real-time physical state of the cloth, achieving dynamic adaptation.

[0091] In some embodiments, the physical state data includes the stress degree and deformation degree of the three-dimensional cloth; the material information of the three-dimensional cloth is determined in advance, and the elastic deformation area and the plastic deformation area on the three-dimensional cloth are determined based on the material information and the visual data; for the elastic deformation area, the coordinates of the corresponding feature points in the three-dimensional point cloud are adjusted according to the stress degree, the deformation degree, and the material information; for the plastic deformation area, the wrinkle convex region and the stacking concave region are determined, and the coordinates of the corresponding feature points in the three-dimensional point cloud are corrected based on the deformation degree of the wrinkle convex region and the stacking concave region. By regionally correcting the three-dimensional point cloud, the three-dimensional point cloud can not only conform to the dynamic changes of the elastic deformation of the cloth, but also accurately restore the fixed shape of the plastic deformation, providing accurate point cloud data basis for subsequent construction of a three-dimensional model that highly matches the real-time physical state of the cloth, and further improving the accuracy of the subsequent center positioning result.

[0092] The stress degree refers to the quantitative value of the external force acting on the cloth during sewing, which can be collected in real time by a pressure sensor or a tension sensor, and the value directly reflects the influence of the external force on the cloth shape; the deformation degree is a quantitative indicator of the shape change of the cloth under the action of the external force, including the stretching length, the contraction amplitude, the wrinkle depth, etc., which is calculated through the displacement change of the feature points in the visual data or the deformation feedback of the physical sensor, and directly reflects the degree of deviation of the cloth shape from the initial state.

[0093] Before the three-dimensional point cloud is corrected, the material information of the three-dimensional cloth needs to be determined in advance, including the key parameters such as the elastic coefficient, plastic threshold and material density of the cloth. When the elastic deformation area and the plastic deformation area of the three-dimensional cloth are determined in combination with the material information and the visual data, the visual data mainly provides the morphological change characteristics of the cloth surface. For example, the elastic deformation area is usually characterized by uniform stretching or contraction of the cloth texture without obvious wrinkle residues, while the plastic deformation area is characterized by texture stacking, fixed wrinkles and inability to recover with the decrease of external force. By comparing the morphological characteristics observed by the visual data with the elastic coefficient and the plastic threshold in the material information, if the deformation degree of a certain area under the action of external force does not exceed the plastic threshold of the material, and there is no fixed wrinkle in the visual, it is determined as the elastic deformation area. If the deformation degree exceeds the plastic threshold, and there are fixed wrinkles or texture stacking in the visual, it is determined as the plastic deformation area.

[0094] When adjusting the coordinates of the corresponding feature points in the three-dimensional point cloud for the elastic deformation area, the adjustment amount needs to be calculated by comprehensively considering the stress degree, deformation degree and material information. First, the corresponding relationship between the external force and the deformation variable is determined according to the elastic coefficient of the material, that is, under the same stress, the smaller the elastic coefficient, the larger the deformation variable. Then, the displacement offset of each feature point in the elastic deformation area due to elastic deformation is calculated by combining the real-time collected stress degree and deformation degree. For example, when the cloth is subjected to tension along the cloth feeding direction, the feature points in the elastic deformation area will produce a displacement matching the stress degree and the elastic coefficient along the tension direction. By superimposing this displacement offset on the coordinates of the original feature points in the three-dimensional point cloud, the coordinates of the point cloud in the elastic deformation area are accurately adjusted, ensuring that the morphology of the point cloud is consistent with the actual morphology of the cloth after elastic deformation.

[0095] For the plastic deformation area, first, the wrinkle protruding area and the stacking recessed area are determined through the depth information of the visual data and the edge detection algorithm. The wrinkle protruding area is characterized by the local morphology of the cloth surface being higher than the surrounding area, and visually presenting a difference in highlight reflection or a depth value higher than that of the surrounding feature points. The stacking recessed area is characterized by the local morphology of the cloth surface being lower than the surrounding area, and visually presenting a shadow or a depth value lower than that of the surrounding feature points. The height and range of the wrinkle protruding area and the depth and area of the stacking recessed area are quantified by combining the deformation degree. When the coordinates of the corresponding feature points in the three-dimensional point cloud are corrected based on these parameters, for the wrinkle protruding area, the z-axis coordinate (height direction) of the corresponding feature point needs to be adjusted upwards to a value matching the height of the wrinkle protruding area, and the x-axis and y-axis coordinates need to be adjusted to fit the contour of the protruding area. For the stacking recessed area, the z-axis coordinate of the corresponding feature point needs to be adjusted downwards to a value matching the depth of the recessed area, and the x-axis and y-axis coordinates need to be adjusted according to the contour of the recessed area. Through this targeted correction, the coordinates of the feature points in the plastic deformation area of the three-dimensional point cloud can accurately reflect the real morphology of the wrinkle protruding and stacking recessed, avoiding the disconnection between the point cloud and the actual cloth morphology caused by plastic deformation.

[0096] The application also provides an AI vision-based three-dimensional fabric center positioning system of a sewing machine, which comprises:

[0097] an input unit configured to acquire environmental interference data, image data and physical state data of the three-dimensional fabric in real time, and perform anti-interference optimization on the image data based on the environmental interference data to obtain vision data;

[0098] a three-dimensional reconstruction unit configured to extract feature data of the three-dimensional fabric from the vision data by using a preset AI recognition algorithm, perform three-dimensional reconstruction based on the feature data to generate three-dimensional point cloud of a fabric surface, and perform dynamic correction on the three-dimensional point cloud based on the physical state data to obtain point cloud data;

[0099] a center positioning unit configured to generate a vision positioning result based on the vision data, construct a three-dimensional model dynamically adapted to a real-time physical form of the fabric in combination with the point cloud data, generate a model positioning result based on the three-dimensional model, and fuse the vision positioning result and the model positioning result to obtain a current center positioning result of the three-dimensional fabric;

[0100] an output unit configured to guide the sewing machine to perform a sewing operation of the three-dimensional fabric based on the current center positioning result.

[0101] The application also provides an intelligent sewing machine configured to implement the AI vision-based three-dimensional fabric center positioning method of any one of the above.

[0102] Those skilled in the art should understand that the modules of the application described above can be realized by a general computing system, which can be concentrated on a single computing system or distributed on a network composed of multiple computing systems, and can be realized by program codes executable by the computing system, so as to be stored in a storage system and executed by the computing system, or be respectively manufactured into integrated circuit modules, or be manufactured into a single integrated circuit module.

[0103] Note that the above are only preferred embodiments of the application and the technical principles applied by the application. Those skilled in the art should understand that the application is not limited to the specific embodiments herein, and those skilled in the art can make various obvious changes, readjustments and substitutions without departing from the protection scope of the application. Therefore, although the application has been described in detail through the above embodiments, the application is not limited to the above embodiments, and can include more other equivalent embodiments without departing from the concept of the application, and the scope of the application is determined by the appended claims.

[0104] The above disclosed are only several specific implementation scenarios of the present application, but the present application is not limited thereto, and any variations that can be thought of by any person skilled in the art shall fall within the protection scope of the present application.

Claims

1. A method for center positioning of three-dimensional fabric in a sewing machine using AI vision, characterized in that, include: Real-time acquisition of environmental interference data, image data and physical state data of three-dimensional fabric; and anti-interference optimization of the image data based on the environmental interference data to obtain visual data. A preset AI recognition algorithm is used to extract feature data of the three-dimensional fabric from the visual data. After three-dimensional reconstruction based on the feature data, a three-dimensional point cloud of the fabric surface is generated. The three-dimensional point cloud is then dynamically corrected based on the physical state data to obtain point cloud data. Based on the visual data, a visual positioning result is generated. A three-dimensional model that dynamically adapts to the real-time physical shape of the fabric is constructed by combining point cloud data. A model positioning result is generated based on the three-dimensional model. The current center positioning result of the three-dimensional fabric is obtained by fusing the visual positioning result and the model positioning result. Based on the current center positioning result, the sewing machine is guided to perform the sewing operation of the three-dimensional fabric.

2. The sewing machine three-dimensional fabric center positioning method according to claim 1, characterized in that, Also includes: By analyzing the temporal variation pattern of the physical state data, the deformation direction and amplitude of the fabric in subsequent periods are predicted, thereby inferring the predicted center positioning result of the three-dimensional fabric. Based on the current center positioning result and the predicted center positioning result, the sewing machine is guided to perform the sewing operation of the three-dimensional fabric.

3. The sewing machine three-dimensional fabric center positioning method according to claim 2, characterized in that, Also includes: Calculate the confidence levels of the current center positioning result and the predicted center positioning result respectively; Based on the real-time sewing speed of the sewing machine and the deformation rate of the three-dimensional fabric, the fusion weights of the current center positioning result and the predicted center positioning result are dynamically allocated; wherein, the fusion weights are positively correlated with their corresponding positioning confidence. The current center positioning result and the predicted center positioning result are weighted and fused based on the fusion weight to generate the final center positioning result, so as to guide the sewing machine to perform sewing operations.

4. The sewing machine three-dimensional fabric center positioning method according to claim 1, characterized in that, The environmental interference data includes ambient light intensity, vibration frequency data, and vibration amplitude data; the image data includes synchronously acquired binocular image pairs. The anti-interference optimization specifically includes: performing adaptive optical correction on the binocular image pair based on ambient light intensity; performing geometric offset compensation on the optically corrected image based on the vibration frequency data and vibration amplitude data, and realizing coordinate correction by calculating the offset compensation value of each feature point in the image.

5. The sewing machine three-dimensional fabric center positioning method according to claim 2, characterized in that, The process of obtaining the predicted center location result specifically includes: Acquire physical state data over a continuous period of time, analyze the changing trend and fluctuation range of the physical state data over that continuous period of time, and determine the dynamic deformation characteristics of the fabric; Historical time-series data, composed of dynamic deformation characteristics and physical state data, are input into a preset time-series prediction model to predict the deformation direction and magnitude of the three-dimensional fabric in subsequent time periods. Based on the current center positioning result of the 3D fabric, combined with the output deformation direction and amplitude, the predicted center positioning result of the 3D fabric in subsequent time periods is calculated.

6. The sewing machine three-dimensional fabric center positioning method according to claim 1, characterized in that, Generate visual localization results based on the visual data, including: The edge feature points and texture feature points of the three-dimensional fabric are extracted from the visual data. Valid feature points are selected by judging whether the spatial distance between the feature points and the surrounding feature points is within a preset range. Based on the three-dimensional coordinates of the valid feature points, the geometric center of all valid feature points is calculated and the geometric center is used as the visual positioning result.

7. The sewing machine three-dimensional fabric center positioning method according to claim 1, characterized in that, Based on the aforementioned 3D model, model localization results are generated, including: The symmetry plane and top surface features of the model are extracted from the three-dimensional model; the intersection line of the symmetry plane and the geometric center of the top surface are calculated, and the intersection line and the geometric center of the top surface are weighted and fused in combination with the material density distribution characteristics of the three-dimensional model; the fused coordinates are used as the model positioning result; wherein, the features corresponding to the region with higher material density have a higher weight in the weighted fusion than the features corresponding to the region with lower density.

8. The sewing machine three-dimensional fabric center positioning method according to claim 1, characterized in that, The process of constructing the three-dimensional model includes: The point cloud data is preprocessed to retain the effective point cloud representing the surface morphology of the fabric; Reconstructing an initial 3D model of the solid cloth based on effective point cloud; The initial 3D model is dynamically adjusted based on real-time acquired physical state data. Specifically, when the physical state data shows that the fabric is stretched, the size of the corresponding area of ​​the model is adjusted along the stretching direction. When the physical state data shows that the fabric has wrinkles, a protrusion or depression structure matching the wrinkle shape is generated in the corresponding area of ​​the model. At preset intervals, the 3D model is readjusted based on the updated point cloud data and physical state data to ensure that the shape of the 3D model always matches the real-time physical shape of the fabric, thus achieving dynamic adaptation.

9. A three-dimensional fabric center positioning system for a sewing machine using AI vision, characterized in that, include: The input unit is used to acquire environmental interference data, image data and physical state data of the three-dimensional fabric in real time, and to perform anti-interference optimization on the image data based on the environmental interference data to obtain visual data. The three-dimensional reconstruction unit is used to extract feature data of the three-dimensional fabric from the visual data using a preset AI recognition algorithm, generate a three-dimensional point cloud of the fabric surface after three-dimensional reconstruction based on the feature data, and dynamically correct the three-dimensional point cloud based on the physical state data to obtain point cloud data. The central positioning unit is used to generate a visual positioning result based on the visual data, construct a three-dimensional model that dynamically adapts to the real-time physical shape of the fabric by combining point cloud data, generate a model positioning result based on the three-dimensional model, and fuse the visual positioning result and the model positioning result to obtain the current central positioning result of the three-dimensional fabric. The output unit is used to guide the sewing machine to perform the sewing operation of the three-dimensional fabric based on the current center positioning result.

10. An intelligent sewing machine, characterized in that, This method is used to implement the AI ​​vision-based three-dimensional fabric center positioning method for sewing machines as described in any one of claims 1-8.

Images