A Sorting Control Method for Flexible Meat Product Production Lines Based on Machine Vision Features

Abstract

This invention discloses a sorting control method for flexible meat product production lines based on machine vision features, belonging to the field of automated meat product sorting technology. It addresses the problems of difficult separation of oily and sticky meat products, easy breakage during handling, and inability to sort according to internal quality. First, a nonlinear speed difference coefficient is generated through visual feedback, driving the centrifugal and conveying mechanisms to physically discretize the stacked raw materials. Then, 3D point clouds and texture features are mapped to theoretical stiffness and friction coefficients, generating a physical property reference benchmark. Next, real-time slip derivatives are monitored during handling, and the benchmark is used to dynamically correct the torque, locking in a stable clamping state. Finally, based on the ratio of closed width to volume, the compression ratio is calculated, and a sorting vector is constructed to complete the classification and placement, achieving non-destructive separation of materials, adaptive flexible detection, and high-precision sorting based on internal physical properties.

Description

Technical Field

[0001] This invention relates to the field of automated meat product sorting technology, specifically a sorting control method for flexible meat product production lines based on machine vision features. Background Technology

[0002] With the large-scale development of the meat processing industry, the consumer demand for cured products such as sausages and bacon is increasing, and the market's requirements for the diversity of product packaging specifications are constantly rising. In order to adapt to the rapid switching of orders of different specifications, flexible production lines are gradually becoming the industry standard. In the meat product packaging process, sorting and material handling are key processes connecting the front-end processing and the back-end packaging. It is usually necessary to separate stacked raw materials and place them into designated pallets according to specifications.

[0003] However, existing meat product sorting methods have significant technical shortcomings in practical applications. Meat products typically have physical characteristics such as an oily surface, uneven texture, and irregular shape, making them difficult to handle with traditional rigid mechanical grippers. For example, when using equipment with constant clamping force, excessive force can easily cause casing rupture or meat deformation; if vacuum suction cups are used, surface grease often leads to seal failure, causing materials to slip. Furthermore, existing vision sorting systems mostly rely on two-dimensional images for positioning and cannot perceive the overlapping or sticking of materials, resulting in the robotic arm missing or grabbing multiple items when handling stacked materials. This lack of perception of the material's physical state leads to high breakage rates and unstable sorting accuracy when dealing with different batches and meat products of varying textures, failing to meet the high-efficiency operation requirements of flexible production lines. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a sorting control method for flexible meat product production lines based on machine vision features, thus solving the problems mentioned in the background.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a sorting control method for flexible meat product production lines based on machine vision features, comprising the following steps: S1. Acquiring stacked images of raw meat products on a conveyor belt, extracting the pixel area and centroid coordinates of connected components, calculating the Euclidean distance between the centroid coordinates of adjacent connected components, and when the Euclidean distance is less than a separation threshold, generating a nonlinear speed difference coefficient based on the ratio of pixel area to standard area, adjusting the rotational speed of the centrifugal mechanism and the speed of the differential conveyor belt, and outputting meat products to be sorted in a single-layer discrete state; S2. Acquiring 3D point cloud and texture images of the single-layer discrete state meat products, calculating the average curvature modulus of the point cloud and the contrast of the texture gray-level co-occurrence matrix, mapping the contrast to the theoretical friction coefficient, mapping the curvature modulus to the theoretical stiffness value, and generating a data set including the initial clamping point cloud. S3. Based on the physical characteristic reference benchmark, drive the flexible robotic arm to perform gripping. During the clamp closing process, collect fingertip pressure and tangential slip, calculate the derivative of tangential slip with respect to fingertip pressure to obtain the real-time slip derivative. When the real-time slip derivative exceeds the allowable slip ratio, calculate the pressure compensation value through the theoretical friction coefficient to correct the driving torque until the slip derivative returns to zero, locking the stable clamping state; S4. In the stable clamping state, acquire the clamp closing width change and fingertip pressure value. Combined with the volume parameters from 3D point cloud computing, calculate the ratio of closing width change to volume parameters to obtain the compression ratio feature. Construct a sorting vector containing fingertip pressure value and compression ratio feature, and drive the robotic arm to place the meat product into the specification channel matching the sorting vector.

[0006] Further, the specific process of acquiring stacked images of meat product raw materials on the conveyor belt, extracting the pixel area and centroid coordinates of connected components, and calculating the Euclidean distance between the centroid coordinates of adjacent connected components is as follows: The stacked image is converted to the HSV color space and the saturation component is extracted. The saturation component is binarized and morphologically opened using an adaptive thresholding algorithm to obtain the foreground mask. Connected component labeling is performed on the foreground mask, and the zero-order moment of the connected component is counted as the pixel area of ​​the connected component. The centroid coordinates of the connected component are calculated by the ratio of the first-order moment to the zero-order moment of the connected component. A connected component adjacency graph is constructed, and pairs of connected components with edge connections are extracted from the adjacency graph. The straight-line distance between the centroid coordinates of the connected component pairs is calculated as the Euclidean distance.

[0007] Furthermore, when the Euclidean distance is less than the separation threshold, a nonlinear speed difference coefficient is generated based on the ratio of the pixel area to the standard area. The centrifugal mechanism rotation speed and the differential conveyor belt speed are adjusted to output the meat products to be sorted in a single-layer discrete state. The specific process is as follows: The ratio of the pixel area to the preset standard area is calculated to obtain the area ratio. The area ratio is substituted into the acceleration gain lookup table to select the centrifugal tangential acceleration gain and the conveyor belt linear acceleration gain. A differential motion equation is constructed through the centrifugal tangential acceleration gain and the conveyor belt linear acceleration gain to generate a nonlinear speed difference coefficient that characterizes the rate of change of the centrifugal mechanism linear velocity and the differential conveyor belt speed. A step speed adjustment command is sent to the centrifugal mechanism and a hysteresis acceleration command is sent to the differential conveyor belt according to the nonlinear speed difference coefficient. The shear force vector generated by the speed difference between the centrifugal mechanism and the differential conveyor belt separates the meat products until the Euclidean distance of the visual feedback is greater than the separation threshold, which is determined to be a single-layer discrete state.

[0008] Further, the specific process of acquiring 3D point cloud and texture images of single-layer discrete meat products, and calculating the average curvature modulus of the point cloud and the contrast of the texture gray-level co-occurrence matrix is ​​as follows: Voxelization downsampling is performed on the 3D point cloud, the divergence of the point cloud normal vectors within the voxel grid is calculated, and the arithmetic mean of the divergence of the normal vectors of all voxel grids is used as the average curvature modulus of the point cloud; the texture image is converted into a grayscale image, the sliding window size and displacement direction vector are set, and the joint probability distribution of the pixel gray levels within the sliding window is statistically analyzed to construct the gray-level co-occurrence matrix; the weighted sum of the squares of the distances between the element values ​​of the gray-level co-occurrence matrix and the main diagonal is calculated, and the weighted sum is used as the contrast of the texture gray-level co-occurrence matrix reflecting the depth of the surface texture grooves.

[0009] Furthermore, the specific process of mapping contrast to theoretical friction coefficient and curvature modulus to theoretical stiffness value to generate a physical property reference benchmark including initial clamping force and allowable slip ratio is as follows: Input the contrast of the texture gray-level co-occurrence matrix into the texture-friction mapping function to obtain the theoretical friction coefficient corresponding to the oil distribution state on the surface of the meat product; input the average curvature modulus of the point cloud into the deformation-stiffness mapping function to obtain the theoretical stiffness value corresponding to the fullness of the meat product's tissue; calculate the critical anti-slip pressure through the theoretical friction coefficient and the gravity of the meat product, and generate the initial clamping force by superimposing a preset safety margin; calculate the elastic deformation limit depth of the meat product through the theoretical stiffness value, deduce the tangential resistance at the elastic deformation limit depth by combining the theoretical friction coefficient, set the slip change rate corresponding to the tangential resistance as the allowable slip ratio, and combine the initial clamping force and the allowable slip ratio to form a physical property reference benchmark.

[0010] Furthermore, the flexible robotic arm is driven to perform grasping based on the physical characteristic reference benchmark. During the clamp closing process, fingertip pressure and tangential slip are collected, and the real-time slip derivative is calculated by differentiating the tangential slip with respect to the fingertip pressure. The specific process is as follows: the initial clamping force contained in the physical characteristic reference benchmark is converted into a current command to drive the clamp servo motor to perform the closing action, and a pressure sensor data sequence and a visual slip detection data sequence containing high-frequency timestamps are established; the pressure sensor data sequence and the visual slip detection data sequence are time-aligned, and the tangential slip increment and fingertip pressure increment under the same timestamp are extracted. The ratio of the tangential slip increment to the fingertip pressure increment is calculated by the differential quotient algorithm, and the ratio of the tangential slip increment to the fingertip pressure increment is set as the real-time slip derivative.

[0011] Furthermore, when the real-time slip derivative exceeds the allowable slip ratio, the driving torque is corrected by calculating the pressure compensation value using the theoretical friction coefficient until the slip derivative returns to zero, thus locking in a stable clamping state. The specific process is as follows: The difference between the real-time slip derivative and the allowable slip ratio is calculated as the slip deviation. The reciprocal of the theoretical friction coefficient is used as the pressure gain factor. The pressure compensation value is generated by multiplying the slip deviation and the pressure gain factor. The pressure compensation value is converted into an additional torque and superimposed on the current driving torque of the servo motor. The driving gripper fingertip increases positive pressure to suppress tangential slippage. The change in the real-time slip derivative is continuously monitored. When the real-time slip derivative is zero for a consecutive preset sampling period, it is determined that the fingertip and the surface of the meat product have reached static friction equilibrium. The current driving torque remains unchanged, and the stable clamping state is locked.

[0012] Furthermore, under stable clamping conditions, the specific process of obtaining the change in clamp closure width and fingertip pressure value, combined with the volume parameters obtained from 3D point cloud computing, and calculating the ratio of the change in closure width to the volume parameters to obtain the compression ratio feature is as follows: The encoder position value of the servo motor under stable clamping conditions is read, and the difference between the encoder position value and the zero-position value of the clamp under no-load closure is calculated. This difference is converted into a physical distance as the change in clamp closure width. The reading of the pressure sensor under stable clamping conditions is read as the fingertip pressure value. The volume parameters of the meat product generated by the 3D point cloud convex hull algorithm are retrieved, and the ratio of the change in clamp closure width to the volume parameters of the meat product is calculated. This ratio is then set as the compression ratio feature.

[0013] Furthermore, the specific process of constructing a sorting vector containing fingertip pressure value and compression ratio features, and driving the robotic arm to place meat products into the specification channel that matches the sorting vector, is as follows: Establish a two-dimensional feature coordinate system, map the fingertip pressure value to the horizontal axis, map the compression ratio feature to the vertical axis, and combine the horizontal and vertical axes to generate the sorting vector of the current meat product; calculate the Euclidean distance between the sorting vector and multiple preset specification cluster centers, select the specification channel corresponding to the cluster center with the smallest Euclidean distance as the target placement channel, plan the motion trajectory of the robotic arm end from the current position to the target placement channel, drive the robotic arm to run along the motion trajectory, and execute the clamp opening command when it reaches the top of the target placement channel to complete the physical sorting of the meat products.

[0014] The present invention has the following beneficial effects:

[0015] (1) A sorting control method for flexible meat product production lines based on machine vision features achieves non-destructive separation and adaptive gripping prediction of materials. By acquiring stacked images and calculating connected component features, the nonlinear speed difference between the centrifugal mechanism and the conveyor belt is actively adjusted through visual feedback. Physical shear force is used to forcibly separate the adhered meat products, solving the problem of accumulation caused by surface grease adhesion of meat products and ensuring that subsequent processes can handle single-layer discrete materials. At the same time, this method innovatively utilizes 3D point cloud and texture image features to map visual information into theoretical friction coefficient and theoretical stiffness value, which is equivalent to establishing a physical property profile for the material before contact. This mechanism enables the control system to generate a physical property reference benchmark containing initial clamping force and allowable slip ratio in advance according to the surface grease and tissue fullness of the meat products, avoiding the inefficiency or initial impact damage caused by blind trial gripping.

[0016] (2) A sorting control method for flexible meat production lines based on machine vision features achieves dynamic and stable gripping and high-precision sorting based on internal physical properties. During the gripping action, a single closing command is not executed; instead, real-time tactile detection is performed. By monitoring the differential of tangential slip with respect to pressure, precise torque compensation can be achieved at the instant of microscopic slippage, based on the pre-calculated friction coefficient. This ensures stable gripping without breaking the casing, effectively resolving the contradiction between unstable gripping and easy breakage of oily soft materials. Using a robotic arm as a measuring tool, the ratio of the closing width to the volume is calculated under stable gripping conditions to directly reflect the internal density and looseness of the meat products. This sorting vector construction method, which combines visual volume parameters with tactile deformation parameters, enables the system to identify meat quality differences that cannot be distinguished by appearance alone, thereby accurately placing materials into matching specification channels and significantly improving the precision of sorting.

[0017] Of course, any product implementing this invention does not necessarily need to achieve all of the advantages described above at the same time. Attached Figure Description

[0018] Figure 1 This is a flowchart of the sorting control method for flexible meat product production lines based on machine vision features, as described in this invention. Detailed Implementation

[0019] This application's embodiments solve the technical problems of high breakage rate, difficulty in separating stacked and adhered products, and inability to classify meat products according to internal quality during automated sorting by a sorting control method for flexible production lines based on machine vision features.

[0020] The overall approach of the solution in this application embodiment is as follows: First, machine vision is used as the perception front end. By analyzing the connected component features in the image, the degree of material stacking and adhesion is identified, and based on this, the centrifugal and conveying mechanisms are driven to generate a speed difference, physically discretizing the stacked materials. Subsequently, the system utilizes 3D vision and texture analysis technology to non-contactly infer the surface friction and structural stiffness characteristics of the material, providing a physical parameter benchmark for subsequent gripping. During the gripping execution phase, the control system combines force sensors and visual slip detection, dynamically adjusting the driving torque of the robotic arm by calculating the slip derivative in real time, achieving adaptive, flexible, and stable gripping of oily soft materials. Finally, the system uses the comparison between deformation data during stable gripping and visual volume data to calculate the compression ratio characteristics characterizing the internal properties of the material, thereby constructing a multi-dimensional sorting vector to drive the execution mechanism to complete accurate classification and placement based on the physical properties of the material.

[0021] Please see Figure 1This invention provides a technical solution: a sorting control method for flexible meat product production lines based on machine vision features, comprising the following steps: S1. Acquiring stacked images of raw meat products on a conveyor belt, extracting the pixel area and centroid coordinates of connected components, calculating the Euclidean distance between the centroid coordinates of adjacent connected components, and when the Euclidean distance is less than a separation threshold, generating a nonlinear speed difference coefficient based on the ratio of pixel area to standard area, adjusting the rotational speed of the centrifugal mechanism and the speed of the differential conveyor belt, and outputting meat products to be sorted in a single-layer discrete state; S2. Acquiring 3D point cloud and texture images of the single-layer discrete state meat products, calculating the average curvature modulus of the point cloud and the contrast of the texture gray-level co-occurrence matrix, mapping the contrast to the theoretical friction coefficient, mapping the curvature modulus to the theoretical stiffness value, and generating a model including the initial clamping force and allowable friction coefficient. Using the physical properties of the slip ratio as a reference benchmark; S3. Drive the flexible robotic arm to perform gripping based on the physical property reference benchmark. During the clamp closing process, collect the fingertip pressure and tangential slip amount, calculate the derivative of the tangential slip amount with respect to the fingertip pressure to obtain the real-time slip derivative. When the real-time slip derivative exceeds the allowable slip ratio, calculate the pressure compensation value through the theoretical friction coefficient to correct the driving torque until the slip derivative returns to zero, locking the stable clamping state; S4. In the stable clamping state, acquire the change in clamp closing width and the fingertip pressure value. Combined with the volume parameters obtained by 3D point cloud computing, calculate the ratio of the change in closing width to the volume parameters to obtain the compression ratio feature. Construct a sorting vector containing the fingertip pressure value and the compression ratio feature, and drive the robotic arm to place the meat products into the specification channel that matches the sorting vector.

[0022] In this implementation scheme, the core function of step S1 is to actively control the physical separation mechanism using visual feedback to ensure the individualized input for subsequent sorting. The system first identifies the stacking state of the raw materials through image processing, where the area of ​​connected pixels reflects the projected size of the material, and the centroid coordinates are used to locate the material's center. Euclidean distance serves as a quantitative indicator of the degree of adhesion; when the distance between adjacent centroids is less than a set threshold, it indicates physical adhesion or stacking. Based on this, the system generates a nonlinear speed difference coefficient, which is a non-constant rate of change control command used to drive the centrifugal mechanism to generate tangential force and simultaneously drive the differential conveyor belt to generate axial tension. This dynamic speed difference adjustment based on visual feedback utilizes physical shear force to forcibly tear the stacked meat products, achieving a physical transformation from disordered stacking to a single-layer discrete state, solving the technical problem of meat products being difficult to separate by traditional fixed-speed conveyor belts due to surface grease adhesion. Step S2 aims to pre-establish a physical characteristic profile of the meat products through non-contact visual detection, providing a mechanical basis for subsequent sorting actions. The system acquires 3D point cloud and texture images, calculates the average curvature modulus of the point cloud to characterize the curvature and fullness of the material surface, and calculates the contrast of the texture gray-level co-occurrence matrix to quantify the surface oil distribution and roughness. Based on this, the system establishes mapping relationships: mapping contrast to the theoretical friction coefficient, i.e., estimating the anti-slip difficulty during gripping; mapping curvature modulus to the theoretical stiffness value, i.e., estimating the material's resistance to deformation. The generated physical property reference benchmark includes the initial clamping force and allowable slip ratio, essentially customizing a gripping strategy for each non-standardized meat product. This process transforms visual features into physical and mechanical parameters. Step S3 transforms the robotic arm's gripping process into a dynamic detection and verification process of the physical state of the meat product. During gripping, the system monitors in real time the derivative of tangential slip with respect to fingertip pressure, i.e., the real-time slip derivative, which characterizes the instantaneous trend of frictional failure. When the slip trend is detected to exceed the allowable range, the system uses the theoretical friction coefficient predicted in S2 to calculate a precise pressure compensation value and fine-tune the driving torque. This closed-loop control is not simply about gripping objects, but rather, through a detection-feedback-correction process, it finds and locks the critical point (stable gripping state) where the fingertips and the material surface reach static friction equilibrium without damaging the soft meat products. This step ensures that the physical state upon which subsequent sorting is based is real and stable, avoiding feature distortion caused by gripping slippage or excessive compression. Step S4 is the key step in ultimately achieving physical attribute classification. In the locked, stable gripping state, the system acquires the change in the clamp's closing width, which represents the deformation of the meat product under specific pressure. Combining the volume parameters measured by 3D point cloud, the system calculates the ratio of the two to obtain the compression ratio feature. The compression ratio feature is a comprehensive physical indicator that can intuitively reflect the internal tissue density, lean-to-fat ratio, and filling tightness of the meat product (for example, for the same volume, meat products with more fat or loose filling have a higher compression ratio).The system constructs a multi-dimensional sorting vector by combining the feature with the fingertip pressure value, and uses it as the direct basis for classification to drive the robotic arm to place the meat products into the corresponding specification channel.

[0023] Specifically, the process of acquiring stacked images of meat product raw materials on a conveyor belt, extracting the pixel area and centroid coordinates of connected components, and calculating the Euclidean distance between the centroid coordinates of adjacent connected components is as follows: The stacked image is converted to the HSV color space and the saturation component is extracted. The saturation component is binarized and morphologically opened using an adaptive thresholding algorithm to obtain the foreground mask. Connected component labeling is performed on the foreground mask, and the zero-order moment of the connected component is counted as the pixel area of ​​the connected component. The centroid coordinates of the connected component are calculated by the ratio of the first-order moment to the zero-order moment of the connected component. A connected component adjacency graph is constructed, and pairs of connected components with edge connections are extracted from the adjacency graph. The straight-line distance between the centroid coordinates of the connected component pairs is calculated as the Euclidean distance.

[0024] In this implementation scheme, the acquired RGB image is first converted to the Hue-Saturation-Lightness color space, i.e., the HSV space. Since the surface of meat products is rich in oil and has a bright color, there is a significant grayscale difference between the saturation component and the conveyor belt background. Utilizing this physical optical characteristic, an adaptive thresholding algorithm is used to binarize the saturation component, and morphological opening operations are used to remove minute oil reflection noise, thereby accurately obtaining the foreground mask representing the meat product area. Subsequently, to quantify the stacking and adhesion between meat products, the system introduces the concept of image moments to perform connected component analysis on the foreground mask. Image moments are weighted averages of image pixel intensities and can reflect the geometric features of objects. For any k-th connected component in the foreground mask, the system determines the connected component pixel area by calculating its zeroth-order moment and combines it with the first-order moment to calculate its geometric centroid coordinates. The specific centroid coordinate calculation formula is as follows: ;in, Indicates the first The zeroth moment of each connected component, i.e., the pixel area of ​​the connected region; and Represents the horizontal and vertical coordinate indices of image pixels; Indicates the first A binary pixel function for each connected component; the value is 1 if the pixel belongs to the connected component, and 0 otherwise. and They represent the first The x and y coordinates of the centroids of each connected component are calculated. This step simplifies the irregular shape of the meat product into a physically meaningful geometric center point, providing a benchmark for subsequent distance measurement. Based on this, the system constructs a connected component adjacency graph, treating each centroid as a node in the graph, and calculates the Euclidean distance between pairs of connected components connected by edges in the adjacency graph. This distance characterizes the spatial proximity of two meat products, and the calculation formula is as follows: ;in, Indicates the first The connected component and the first Euclidean distance between the centroids of adjacent connected domains; and Indicates the first The x and y coordinates of the centroids of each connected region are used to determine whether the two parts are physically adhered. If this distance is less than a preset separation threshold, the two parts are considered to be physically adhered. The separation threshold here is determined by statistically analyzing the average radius of a single standard meat product laid flat on the conveyor belt and multiplying it by a safety factor, and is used to distinguish between natural proximity and substantial adhesion.

[0025] Specifically, when the Euclidean distance is less than the separation threshold, a nonlinear speed difference coefficient is generated based on the ratio of the pixel area to the standard area. The centrifugal mechanism rotation speed and the differential conveyor belt speed are adjusted to output meat products in a single-layer discrete state. The specific process is as follows: The ratio of the pixel area to the preset standard area is calculated to obtain the area ratio. The area ratio is substituted into the acceleration gain lookup table to select the centrifugal tangential acceleration gain and the conveyor belt linear acceleration gain. A differential motion equation is constructed through the centrifugal tangential acceleration gain and the conveyor belt linear acceleration gain to generate a nonlinear speed difference coefficient that characterizes the rate of change of the centrifugal mechanism linear velocity and the differential conveyor belt speed. A step speed adjustment command is sent to the centrifugal mechanism and a hysteresis acceleration command is sent to the differential conveyor belt according to the nonlinear speed difference coefficient. The shear force vector generated by the speed difference between the centrifugal mechanism and the differential conveyor belt separates the meat products until the Euclidean distance of the visual feedback is greater than the separation threshold, which is then determined to be a single-layer discrete state.

[0026] In this implementation scheme, the system first calculates the ratio of the pixel area of ​​the current connected region to the preset standard area to obtain the area ratio. This ratio reflects the aggregation scale of the adhered meat products; a larger ratio means a greater quantity or mass of adhered meat products, requiring stronger physical shearing force. The system substitutes the area ratio into a preset acceleration gain lookup table and selects suitable centrifugal tangential acceleration gain and conveyor belt linear acceleration gain. To achieve effective separation while protecting the meat products from breakage, the system constructs a differential motion equation to generate a nonlinear speed difference coefficient. This coefficient is not a constant value but a dynamically changing control variable that varies with the degree of adhesion, used to define the severity of the speed difference between the centrifugal mechanism and the conveyor belt. The calculation formula for the nonlinear speed difference coefficient is as follows: ;in, The nonlinear velocity difference coefficient represents the shear strength trend applied to the adhesive interface; This represents the area ratio, which is the ratio of the area of ​​a connected component pixel to the standard area. This represents the centrifugal tangential acceleration gain, used to control the response amplitude of the centrifugal mechanism in generating the tangential ejection force; This represents the linear acceleration gain of the conveyor belt, used to control the response amplitude of the axial tension generated by the differential conveyor belt. This represents the hysteresis response factor, used to set the time delay between the conveyor belt acceleration and the centrifugal mechanism acceleration. The introduction of this factor creates the instantaneous peak speed difference. is the base of the natural logarithm. This formula utilizes the logarithmic function. Its characteristics ensure sufficient tangential force for high-magnification aggregates, while preventing centrifugal force runaway due to excessive gain; utilizing the exponential decay term The timing of the intervention of the linear tension force is controlled to form a composite force field of "first throwing, then pulling". Finally, based on the calculated nonlinear speed difference coefficient, the control system sends a step speed regulation command to the centrifugal mechanism to make it rotate rapidly and generate tangential shear force, while simultaneously sending a delayed acceleration command to the differential conveyor belt to generate axial tensile force. After the synthesis of these two force vectors with orthogonal directions and time difference, a severe shear tearing effect is formed at the adhesion interface of the meat products, until the Euclidean distance fed back by the vision system is greater than the separation threshold, indicating that the meat products have been successfully separated into a single-layer discrete state.

[0027] Specifically, the process of acquiring 3D point cloud and texture images of a single-layer discrete state meat product, and calculating the average curvature modulus of the point cloud and the contrast of the texture gray-level co-occurrence matrix is ​​as follows: Voxelization downsampling is performed on the 3D point cloud; the divergence of the point cloud normal vectors within the voxel grid is calculated; the arithmetic mean of the divergence of all voxel grid normal vectors is used as the average curvature modulus of the point cloud; the texture image is converted to a grayscale image; the sliding window size and displacement direction vector are set; the joint probability distribution of pixel gray levels within the sliding window is statistically analyzed to construct the gray-level co-occurrence matrix; the weighted sum of the squares of the distances between the elements of the gray-level co-occurrence matrix and the main diagonal is calculated; this weighted sum is used as the contrast of the texture gray-level co-occurrence matrix, reflecting the depth of the surface texture grooves.

[0028] In this implementation scheme, the 3D point cloud data acquired by the depth camera is processed first. Since the raw point cloud data is massive and contains spatial noise, direct computation is inefficient. Therefore, voxelization downsampling is performed first, dividing the continuous 3D space into tiny cubic meshes, i.e., voxels. The centroid of all points within a voxel is used to replace the voxel, reducing the data dimensionality while preserving macroscopic geometric features. Subsequently, to quantify the surface fullness and curvature characteristics of meat products, the system calculates the divergence of the point cloud normal vectors within each voxel mesh. The normal vector represents the surface orientation, and its divergence reflects the degree of curvature change in a local area. The system calculates the arithmetic mean of the divergence of the normal vectors of all voxel meshes, defining it as the average curvature modulus of the point cloud. The formula for calculating this modulus is as follows: ;in, This represents the average curvature modulus of the point cloud. The larger the value, the more uneven or overly full the surface of the meat product. Indicates the total number of effective voxel grids; Indicates the index of the current voxel grid; Indicates the first Principal normal vectors of individual element meshes; Indicates the first The spatial neighborhood set of individual element grids; Indicates the number of neighbors within the neighborhood. The normal vector of an individual element; This represents the unit distance vector between the centers of two voxels; The divergence operator is approximated here using a discrete difference form. Simultaneously, texture images are processed to analyze surface friction characteristics. After converting the color texture image to a grayscale image, a gray-level co-occurrence matrix (GLCM) is constructed by setting a sliding window of a specific size and a displacement direction vector. The GLCM describes the probability of grayscale occurrence for pixel pairs with specific distances and orientations in the image, effectively characterizing texture roughness. The system calculates the contrast feature of the GLCM, which reflects the severity of local grayscale changes in the image. For meat products with a high surface fat content and smoothness, the texture is blurred and the contrast is low; while for meat products with a dry surface or many casing wrinkles, the texture has deep grooves and the contrast is high. The formula for calculating the contrast of the texture GLCM is as follows: ;in, Represents the contrast of the texture gray-level co-occurrence matrix; This represents the number of gray levels in the image, typically 256. and These represent the grayscale values ​​of the two pixels in a pixel pair; This indicates that within the sliding window, along the direction of displacement... And the distance is The gray levels of the two pixels are respectively and The joint probability distribution value; As a weighting term, this term causes the contribution of high grayscale difference elements far from the main diagonal to the contrast to increase quadratically, thereby sensitively capturing the depth of grooves in surface texture.

[0029] Specifically, the process of mapping contrast to theoretical friction coefficient and curvature modulus to theoretical stiffness value to generate a physical property reference benchmark including initial clamping force and allowable slip ratio is as follows: Input the contrast of the texture gray-level co-occurrence matrix into the texture-friction mapping function to obtain the theoretical friction coefficient corresponding to the oil distribution state on the surface of the meat product; input the average curvature modulus of the point cloud into the deformation-stiffness mapping function to obtain the theoretical stiffness value corresponding to the fullness of the meat product's tissue; calculate the critical anti-slip pressure by using the theoretical friction coefficient and the gravity of the meat product, and generate the initial clamping force by superimposing a preset safety margin; calculate the elastic deformation limit depth of the meat product by using the theoretical stiffness value, deduce the tangential resistance at the elastic deformation limit depth by combining the theoretical friction coefficient, set the slip change rate corresponding to the tangential resistance as the allowable slip ratio, and combine the initial clamping force and the allowable slip ratio to form a physical property reference benchmark.

[0030] In this implementation scheme, the visually extracted features are mapped to physical and mechanical parameters, thereby generating a physical characteristic reference benchmark to guide the robotic arm's grasping. First, using a pre-calibrated mapping function, the contrast of the texture gray-level co-occurrence matrix is ​​converted into the theoretical coefficient of friction. This process is based on the tribological principle that the deeper the surface microtexture, the stronger the mechanical clamping action and the higher the coefficient of friction; conversely, the smooth surface covered with grease has a lower coefficient of friction. Simultaneously, the average curvature modulus of the point cloud is mapped to a theoretical stiffness value, which reflects the resistance of the meat product to pressure. Areas with high curvature modulus typically correspond to areas with dense internal filling or hardened structures. Based on these two derived physical parameters, the system first calculates the initial clamping force to ensure that the material does not slip or damage itself during grasping. The formula for calculating the initial clamping force is as follows: ;in, This indicates the initial clamping force of the end effector gripper of the robotic arm; This represents the estimated weight of the meat product itself, which is obtained by multiplying the volume parameter by the preset density. This represents the preset basic safety margin factor, used to cover the effects of inertial forces during dynamic handling; This represents the theoretical friction coefficient obtained from the contrast mapping. As a denominator, it reflects the inverse relationship that the smaller the friction coefficient, the greater the clamping force required. Indicates the stiffness compensation coefficient; This represents the theoretical stiffness value obtained from the curvature modulus mapping; The first value represents the reference stiffness of standard meat products, while the second part is used to fine-tune the basic clamping force for materials that are too hard or too soft. Next, the system determines the allowable slip ratio, a core threshold in the dynamic force control process. This threshold is not a fixed value but is derived from the elastic limit of the meat product. The system calculates the maximum elastic deformation depth that the meat product can withstand without plastic failure at the theoretical stiffness value, and combines this with the theoretical friction coefficient to calculate the maximum tangential resistance that the interface can provide at this depth. Finally, the slip change rate corresponding to this resistance is set as the allowable slip ratio. The formula for calculating the allowable slip ratio is as follows: ;in, This represents the allowable slip ratio, which is the maximum upper limit of the real-time slip derivative allowed by the system. This parameter represents the elastic deformation limit depth of meat products. It is obtained by fitting destructive compression test data of meat products with different stiffnesses. It represents the normal reaction force at the ultimate depth; This represents the damping coefficient of the gripping system, used to convert the force dimension into the time rate of change dimension of the slip derivative. Through the above calculations, the initial gripping force and the allowable slip ratio are combined to form a physical characteristic reference benchmark. This benchmark provides a specific force control target for the subsequent S3 step: the robotic arm will close with the initial gripping force and dynamically maintain this position within a range where the real-time slip derivative is lower than the allowable slip ratio; if it exceeds this range, pressure compensation is triggered, thereby achieving adaptive flexible sorting based on the material's physical properties.

[0031] Specifically, the process of driving a flexible robotic arm to perform grasping based on a physical characteristic reference benchmark, collecting fingertip pressure and tangential slip during the clamp closing process, and calculating the real-time slip derivative by differentiating the tangential slip with respect to the fingertip pressure is as follows: The initial clamping force contained in the physical characteristic reference benchmark is converted into a current command to drive the clamp servo motor to perform the closing action, and a pressure sensor data sequence and a visual slip detection data sequence containing high-frequency timestamps are established; the pressure sensor data sequence and the visual slip detection data sequence are time-aligned, and the tangential slip increment and the fingertip pressure increment at the same timestamp are extracted. The ratio of the tangential slip increment to the fingertip pressure increment is calculated using a differential quotient algorithm, and the ratio of the tangential slip increment to the fingertip pressure increment is set as the real-time slip derivative.

[0032] In this implementation scheme, the physical characteristic reference base generated in the previous steps is first invoked, and the initial clamping force contained therein is converted into a current control command to drive the clamp servo motor to close. Simultaneously with this physical action, the system initiates a multi-sensor data synchronous acquisition mechanism. Since pressure sensors (tactile) and visual slip detection systems (visual) often have differences in sampling frequency and transmission delay, the system establishes a unified high-frequency timestamp index to time-align the data sequences from the two heterogeneous data sources, ensuring that the fingertip pressure value at the same moment accurately corresponds to the current tangential slip amount. To quantify whether the current grasping state is becoming unstable, the system performs real-time slip derivative calculation based on the differential quotient algorithm. This calculation does not simply compare the absolute value of the slip amount, but rather reflects the sensitivity of the slip state to pressure changes by calculating the derivative of the tangential slip amount with respect to the fingertip pressure. The formula for calculating the real-time slip derivative is as follows: ;in, Indicates in The real-time slip derivative at time t is a parameter that characterizes the instantaneous trend of frictional failure. and These represent the cumulative tangential slip at the current time and the previous sampling time, respectively, and the difference between the two is the tangential slip increment. and These represent the readings of the fingertip pressure sensor at the current time and the previous sampling time, respectively. The difference between the two is the fingertip pressure increment. This represents the sampling time interval of the sensor data. The physical meaning of this formula is: if A positive and large value indicates that although the pressure is increasing, the slippage is still growing rapidly, suggesting that the current clamping force is far from reaching the critical value required to overcome the surface greasiness; if When the value approaches zero, it indicates that slippage no longer occurs with changes in time or pressure, and the contact surface has entered a state of static friction. This differential-based calculation method can capture the dynamic trend of slippage earlier than simple threshold judgment.

[0033] Specifically, when the real-time slip derivative exceeds the allowable slip ratio, the driving torque is corrected by calculating the pressure compensation value using the theoretical friction coefficient until the slip derivative returns to zero, thus locking in a stable clamping state. The specific process is as follows: The difference between the real-time slip derivative and the allowable slip ratio is calculated as the slip deviation. The reciprocal of the theoretical friction coefficient is used as the pressure gain factor. The pressure compensation value is generated by multiplying the slip deviation and the pressure gain factor. The pressure compensation value is converted into an additional torque and superimposed on the current driving torque of the servo motor. The driving gripper fingertip increases positive pressure to suppress tangential slip. The change in the real-time slip derivative is continuously monitored. When the real-time slip derivative is zero for a consecutive preset sampling period, it is determined that the fingertip and the surface of the meat product have reached static friction equilibrium. The current driving torque is kept unchanged, and the stable clamping state is locked.

[0034] In this implementation scheme, for unstable gripping states where the real-time slip derivative exceeds the allowable range, the system performs closed-loop torque correction based on a physical model. When the real-time slip derivative is detected to be greater than the allowable slip ratio set in the previous step, it indicates that the current frictional force is insufficient to resist the gravity or inertial force of the meat product. At this time, the system does not blindly increase the pressure, but calculates an accurate pressure compensation value based on the theoretical friction coefficient to achieve anti-slip with minimal additional force, avoiding excessive compression and damage to the meat product. The formula for calculating the pressure compensation value is as follows: ;in, This indicates the pressure compensation value that needs to be applied in the current control cycle; The slip derivative is calculated in real time; The allowable slip ratio generated in the preceding steps is used as a reference target value for the control system; This constitutes the slip deviation, reflecting the gap between the current degree of instability and the safe state; Let be the theoretical friction coefficient obtained from texture feature mapping, and its reciprocal. As a pressure gain factor, it reflects the physical law that the lower the friction coefficient (the greasier it is), the greater the force feedback gain should be for slippage deviation. The response stiffness coefficient set for the system is used to adjust the response speed of torque correction and prevent overshoot oscillation. The system converts the calculated pressure compensation value into an additional current and adds it to the current drive current of the servo motor, driving the gripper fingertip to apply additional positive pressure. As the positive pressure increases, the maximum static friction force on the contact surface increases, tangential slippage is gradually suppressed, and the real-time slip derivative decreases accordingly. The system continuously monitors this parameter, and when the real-time slip derivative remains zero for a consecutive preset sampling period (e.g., 5 consecutive periods, i.e., within tens of milliseconds), it is determined that the fingertip and the surface of the meat product have reached complete static friction equilibrium. The system locks the current drive torque, no longer increases the pressure, and marks this state as a stable gripping state. This process ensures that the robotic arm not only grasps the material, but also grasps the material in a critically stable state where it just does not slip, providing a reliable physical premise for subsequent deformation-based precise sorting.

[0035] Specifically, under a stable clamping state, the process of obtaining the change in clamp closure width and fingertip pressure value, and combining it with the volume parameters obtained from 3D point cloud computing, to calculate the ratio of the change in closure width to the volume parameters to obtain the compression ratio feature is as follows: The encoder position value of the servo motor under stable clamping state is read, and the difference between the encoder position value and the zero-position value of the clamp under no-load closure is calculated. This difference is converted into a physical distance as the change in clamp closure width. The reading of the pressure sensor under stable clamping state is read as the fingertip pressure value. The volume parameters of the meat product generated by the 3D point cloud convex hull algorithm are retrieved, and the ratio of the change in clamp closure width to the volume parameters of the meat product is calculated. This ratio is then set as the compression ratio feature.

[0036] In this implementation scheme, after the robotic arm reaches a stable gripping state, the system performs feature extraction based on multi-source data fusion. First, to accurately obtain the geometric deformation of the meat product under pressure, the system reads the current position value of the servo motor encoder. The encoder is a rotation sensor mounted on the motor shaft, capable of converting angular displacement into digital pulse signals. The system calculates the difference between this value and the zero-position value when the gripper is fully closed under no-load conditions. Using the reduction ratio parameter of the transmission mechanism, this pulse difference is converted into the actual physical distance moved by the gripper fingertip, i.e., the change in gripper closure width. This variable directly reflects the radial compression depth of the meat product under the current pressure. Next, the system retrieves the meat product volume parameters generated in the previous step (S2) using the 3D point cloud convex hull algorithm. The convex hull algorithm is a geometric algorithm that calculates the minimum convex polyhedron containing all data points, capable of accurately estimating the solid volume of irregular meat products. To eliminate the influence of individual meat product size differences on deformation assessment, the system calculates the ratio of the change in gripper closure width to the meat product volume parameters, obtaining a normalized compression ratio feature. The calculation formula for this feature is as follows: ;in, The compression ratio is a dimensionless physical index used to characterize the firmness and filling density of the internal structure of meat products. This represents the change in the clamp closure width, which is the absolute deformation of the meat product in the direction of pressure. This represents the convex hull volume parameter of the meat product obtained from 3D point cloud computing; This represents a dimensional correction coefficient used to unify the magnitudes of deformation and volume to the same order of magnitude. This coefficient is determined through linear regression analysis of destructive compression test data from standard samples. By applying this formula, the system can distinguish between large, soft meat products and small, hard meat products, achieving a deep quantification of their internal physical properties. Simultaneously, the system directly reads the steady-state reading from the pressure sensor as the final fingertip pressure value. This value, as another dimension of physical characteristic, characterizes the frictional properties and compressive strength of the meat product surface.

[0037] Specifically, the process of constructing a sorting vector containing fingertip pressure value and compression ratio features, and driving the robotic arm to place meat products into the specification channel that matches the sorting vector, is as follows: Establish a two-dimensional feature coordinate system, map the fingertip pressure value to the horizontal axis, map the compression ratio feature to the vertical axis, and combine the horizontal and vertical axes to generate the sorting vector of the current meat product; calculate the Euclidean distance between the sorting vector and multiple preset specification cluster centers, select the specification channel corresponding to the cluster center with the smallest Euclidean distance as the target placement channel, plan the motion trajectory of the robotic arm end from the current position to the target placement channel, drive the robotic arm to run along the motion trajectory, and execute the clamp opening command when it reaches the top of the target placement channel to complete the physical sorting of the meat products.

[0038] In this implementation scheme, the system constructs a sorting vector and performs spatial matching. To comprehensively utilize tactile feedback (pressure) and deformation feedback (compression ratio) for classification, the system establishes a two-dimensional feature coordinate system, mapping the final fingertip pressure value and compression ratio feature to horizontal and vertical coordinates respectively, and combining them to generate the sorting vector for the current meat product. Since pressure value and compression ratio have different physical dimensions, direct combination may lead to an excessively large weight for one feature; therefore, normalization processing is required before mapping. The system calculates the Euclidean distance between this sorting vector and multiple preset cluster centers for different specifications. The cluster centers are the feature mean points obtained by performing K-means clustering analysis on a large amount of historical sorting data of meat products of known specifications (such as premium, grade 1, grade 2, or different fat-to-lean ratios), representing the standard physical shape of each specification. The formula for calculating the Euclidean distance is as follows: ;in, This indicates the sorting vector of the current meat product and the... The weighted Euclidean distance between the cluster centers of each specification; This is the index number for the specification channel; This represents the final fingertip pressure value. For the first The pressure characteristic components of each size cluster center; This is a characteristic of compression ratio; For the first Compression ratio feature components of each size cluster center; and These are the pressure feature weighting coefficient and the compression ratio feature weighting coefficient, respectively. These two coefficients are determined by analyzing the contribution of each feature to the classification result (such as information gain rate) in historical misclassification data, and are used to balance the importance of different physical attributes in classification decisions. The system iterates through all preset cluster centers and selects the channel corresponding to the cluster center with the smallest calculated Euclidean distance as the target delivery channel. Mathematically, this implements the function of a minimum distance classifier; physically, it classifies meat products with unknown characteristics into the standard group closest to their physical attributes. Finally, the control system uses an inverse kinematics algorithm to plan the smooth motion trajectory of the robotic arm's end effector from the current gripping position to above the target delivery channel, driving the robotic arm along this trajectory. When the end effector reaches the target position, it executes the gripper opening command to release the meat product, completing a fully automated closed-loop operation from physical feature perception to physical spatial distribution.

[0039] In summary, this application has at least the following effects:

[0040] This machine vision-based sorting control method for flexible meat production lines utilizes machine vision feedback to drive a centrifugal mechanism and conveyor belt to generate a nonlinear speed difference. It leverages physical shear force to achieve non-destructive discretization of highly adhesive stacked meat products, solving the problem of separating oily, soft materials that is difficult to separate using traditional methods. Furthermore, this method innovatively maps 3D point clouds and texture features to theoretical stiffness and friction coefficients, establishing a physical property reference benchmark. During the gripping process, it monitors real-time slip derivatives for dynamic torque compensation, achieving adaptive flexible and stable detection of non-standardized meat products. Further, it calculates compression ratio features using the change in closed width and volume parameters under stable gripping conditions, constructing a multi-dimensional sorting vector. This enables high-precision classification based on the internal density and looseness of meat products, effectively improving the automated sorting efficiency and quality of flexible production lines for multi-specification and multi-quality meat products.

[0041] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0042] This invention is described with reference to flowchart illustrations and / or block diagrams of systems, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0043] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0044] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0045] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0046] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method of flexible production line sorting control of meat products based on machine vision features, characterized in that, The method comprises the following steps: S1. Collecting the stacked image of the meat product raw material on the conveying belt, extracting the connected domain pixel area and the centroid coordinates, calculating the Euclidean distance of the centroid coordinates of adjacent connected domains, when the Euclidean distance is less than the separation threshold, generating a nonlinear speed difference coefficient according to the ratio of the pixel area to the standard area, adjusting the rotating speed of the centrifugal mechanism and the speed of the differential conveying belt, and outputting the meat product in a single-layer discrete state; S2. Obtain the 3D point cloud and texture image of the single-layer discrete state meat product, calculate the point cloud average curvature modulus and texture gray level co-occurrence matrix contrast, map the contrast to the theoretical friction coefficient, map the curvature modulus to the theoretical stiffness value, and generate a physical property reference benchmark containing the initial clamping force and the allowable slip rate; S3. According to the physical property reference benchmark, drive the flexible mechanical arm to perform grabbing, collect the fingertip pressure and tangential slip amount during the closing of the clamp, calculate the differential of the tangential slip amount to the fingertip pressure to obtain the real-time slip derivative, when the real-time slip derivative exceeds the allowable slip rate, calculate the pressure compensation value through the theoretical friction coefficient to correct the driving torque until the slip derivative is zero, and lock the stable clamping state; S4. In the stable clamping state, obtain the change amount of the closing width of the clamp and the fingertip pressure value, combine the volume parameter calculated from the 3D point cloud, calculate the compression ratio characteristic by the ratio of the closing width change amount to the volume parameter, construct a sorting vector containing the fingertip pressure value and the compression ratio characteristic, and drive the mechanical arm to put the meat product into a specification channel matched with the sorting vector.

2. The machine vision feature based meat flexible production line sortation control method of claim 1, wherein: The specific process of collecting the stacked image of the meat product raw material on the conveying belt, extracting the connected domain pixel area and the centroid coordinates, and calculating the Euclidean distance of the centroid coordinates of adjacent connected domains is as follows: Convert the stacked image to HSV color space and extract the saturation component, perform binaryzation segmentation and morphological opening operation on the saturation component through adaptive threshold algorithm, and obtain the foreground mask; Perform connected component labeling on the foreground mask, count the connected component zeroth moment as the connected domain pixel area, calculate the connected domain centroid coordinates through the ratio of the connected component first moment to the zeroth moment, construct a connected domain adjacency graph, extract the connected domain pairs connected by edges in the adjacency graph, and calculate the straight line distance between the centroid coordinates of the connected domain pairs as the Euclidean distance.

3. The machine vision feature based meat product flexible line sortation control method of claim 2, wherein: When the Euclidean distance is less than the separation threshold, a nonlinear speed difference coefficient is generated according to the ratio of the pixel area to the standard area, the rotating speed of the centrifugal mechanism and the speed of the differential conveying belt are adjusted, and the meat product in a single-layer discrete state is outputted. The specific process is as follows: Calculate the area ratio of the pixel area to the preset standard area to obtain the area ratio, and substitute the area ratio into the acceleration gain lookup table to select the centrifugal tangential acceleration gain and the conveying belt linear acceleration gain; Construct a difference motion equation through the centrifugal tangential acceleration gain and the conveying belt linear acceleration gain, and generate a nonlinear speed difference coefficient representing the change rate of the centrifugal mechanism linear speed and the differential conveying belt speed; According to the nonlinear speed difference coefficient, a step speed command is sent to the centrifugal mechanism, and a lag acceleration command is sent to the differential conveying belt. The shear force vector generated by the speed difference between the centrifugal mechanism and the differential conveying belt separates the meat products until the Euclidean distance of the visual feedback is greater than the separation threshold, and the single-layer discrete state is determined.

4. The machine vision feature based meat flexible production line sortation control method of claim 1, wherein: The specific process of obtaining the 3D point cloud and texture image of the single-layer discrete state meat product and calculating the point cloud average curvature modulus and texture gray level co-occurrence matrix contrast is as follows: Perform voxelization downsampling on the 3D point cloud, calculate the divergence of the point cloud normal vector in the voxel grid, and calculate the arithmetic mean of the normal vector divergence of all voxel grids as the point cloud average curvature modulus; Convert the texture image to a grayscale image, set the sliding window size and displacement direction vector, and calculate the joint probability distribution of the pixel gray level in the sliding window to construct the gray level co-occurrence matrix; Calculate the square weighted sum of the element values and the main diagonal distance of the gray level co-occurrence matrix, and take the weighted sum as the texture gray level co-occurrence matrix contrast reflecting the surface texture ditch depth.

5. The machine vision feature based meat product flexible line sortation control method of claim 4, wherein: The specific process of mapping the contrast to the theoretical friction coefficient and mapping the curvature modulus to the theoretical stiffness value to generate a physical property reference benchmark containing the initial clamping force and the allowable slip rate is as follows: Input the texture gray level co-occurrence matrix contrast into the texture-friction force mapping function to obtain the theoretical friction coefficient corresponding to the grease distribution state of the meat product surface; Input the point cloud average curvature modulus into the deformation-stiffness mapping function to obtain the theoretical stiffness value corresponding to the meat product tissue fullness, calculate the critical anti-skid pressure through the theoretical friction coefficient and the gravity of the meat product, and superimpose a preset safety margin to generate the initial clamping force; Calculate the elastic deformation limit depth of the meat product through the theoretical stiffness value, derive the tangential resistance at the elastic deformation limit depth combined with the theoretical friction coefficient, set the slip change rate corresponding to the tangential resistance as the allowable slip rate, and combine the initial clamping force and the allowable slip rate to form the physical property reference benchmark.

6. The machine vision feature based meat flexible production line sortation control method of claim 1, wherein: According to the physical property reference benchmark, drive the flexible mechanical arm to perform grabbing, and collect the fingertip pressure and tangential slip amount during the closing process of the clamp to calculate the differential of the tangential slip amount to the fingertip pressure to obtain the specific process of the real-time slip derivative as follows: Convert the initial clamping force contained in the physical property reference benchmark to a current command to drive the clamp servo motor to perform closing action, and establish a pressure sensor data sequence and a visual slip detection data sequence containing high-frequency time stamps; Perform time alignment on the pressure sensor data sequence and the visual slip detection data sequence, extract the tangential slip increment and the fingertip pressure increment under the same time stamp, calculate the ratio of the tangential slip increment and the fingertip pressure increment through the difference quotient algorithm, and set the ratio of the tangential slip increment and the fingertip pressure increment as the real-time slip derivative.

7. The machine vision feature based meat product flexible line sortation control method of claim 6, wherein: When the real-time slip derivative exceeds the allowable slip rate, calculate the pressure compensation value through the theoretical friction coefficient to correct the driving torque until the slip derivative is zero, and lock the stable clamping state as follows: Calculate the difference between the real-time slip derivative and the allowable slip rate as the slip deviation, take the inverse of the theoretical friction coefficient as the pressure gain factor, and generate the pressure compensation value through the product of the slip deviation and the pressure gain factor; The pressure compensation value is converted into an additional torque superimposed on the current driving torque of the servo motor, the clamp finger increases the normal pressure to suppress the tangential slip, the change of the real-time slip derivative is continuously monitored, when the real-time slip derivatives of the continuous preset sampling period are all zero, it is determined that the finger and the meat product surface reach the static friction balance, the current driving torque is kept unchanged, and the stable clamping state is locked.

8. The machine vision feature based meat flexible production line sortation control method of claim 1, wherein: In the stable clamping state, the clamp closing width change and the finger pressure value are obtained, the volume parameter calculated by the 3D point cloud is combined, the specific process of calculating the ratio of the closing width change and the volume parameter to obtain the compression ratio characteristic is as follows: Read the encoder position value of the servo motor in the stable clamping state, calculate the difference between the encoder position value and the clamp no-load closing zero value, convert the difference into a physical distance as the clamp closing width change; Read the reading of the pressure sensor in the stable clamping state as the finger pressure value, call the meat product volume parameter generated by the 3D point cloud convex hull algorithm, calculate the ratio of the clamp closing width change and the meat product volume parameter, and set the ratio as the compression ratio characteristic.

9. The machine vision feature based meat product flexible line sortation control method of claim 8, wherein: The specific process of constructing a sorting vector containing the finger pressure value and the compression ratio characteristic, and driving the mechanical arm to put the meat product into the specification channel matched with the sorting vector is as follows: A two-dimensional feature coordinate system is established, the finger pressure value is mapped as the horizontal coordinate, the compression ratio characteristic is mapped as the vertical coordinate, and the horizontal coordinate and the vertical coordinate are combined to generate the sorting vector of the current meat product; Calculate the Euclidean distance between the sorting vector and the preset multiple specification cluster centers, select the specification channel corresponding to the cluster center with the smallest Euclidean distance as the target delivery channel, plan the motion trajectory of the mechanical arm from the current position to the target delivery channel, drive the mechanical arm to run along the motion trajectory, and execute the clamp opening instruction when reaching above the target delivery channel to complete the physical sorting of the meat product.

Images