A Non-Contact Method and System for Estimating the Body Size and Weight of Goats Based on Enhanced Point Cloud Segmentation

By enhancing point cloud segmentation technology, the accuracy and stability issues of goat weight and body size measurement have been solved, enabling contactless measurement and making it suitable for individual monitoring and group management in large-scale farms.

CN123118735APending Publication Date: 2026-11-17SOUTHWEST UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611629045.4
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-10-21
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing technologies for measuring the weight and body size of goats are labor-intensive and prone to causing stress. Two-dimensional images lack three-dimensional geometric information, and existing point cloud methods are insufficient for the body shape characteristics of goats. Direct application of these methods can easily lead to measurement deviations and inaccuracies.

Method used

An augmented point cloud segmentation method is adopted. Point-by-point labels are obtained through the GKA augmented point cloud segmentation network. Body axis, width axis and height axis are established. Combined with multi-section extraction and curve fitting, chest circumference, abdominal circumference, chest width, hip width, body length and body height are calculated. Weight is estimated using three-dimensional morphological information and confidence assessment is performed.

Benefits of technology

It enables stable measurement of goat body size and weight without human contact, improving the accuracy and stability of measurements, adapting to the measurement needs of different breeds and growth stages, and reducing labor intensity and stress risk.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123118735A_ABST
    Figure CN123118735A_ABST
Patent Text Reader

Abstract

This invention discloses a non-contact method and system for estimating the body size and weight of goats based on enhanced point cloud segmentation, belonging to the field of intelligent livestock 3D vision measurement technology. The method includes: acquiring and preprocessing a 3D point cloud of a goat; obtaining eight body part labels using a GKA enhanced point cloud segmentation network; establishing body axis, width axis, and height axis; measuring chest circumference, abdominal circumference, chest width, and rump width through multi-section B-spline fitting and robust statistics; obtaining body length using the arc length of the 3D B-spline curve at key back points; obtaining body height from the horizontal support surface of the limbs; constructing a morphological feature vector by combining volume and surface area; regressing and outputting the estimated weight value; and having the confidence level assessed by a quality control module. The system includes modules for acquisition, preprocessing, segmentation, body size measurement, volume and surface area calculation, weight estimation, quality control, and result output. This invention eliminates the need for manual contact, reducing labor intensity and stress, and improving the stability of body size measurement and the accuracy of weight estimation under point cloud conditions.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention belongs to the field of intelligent livestock three-dimensional visual measurement technology, and in particular relates to a non-contact method and system for estimating the body size and weight of goats based on enhanced point cloud segmentation. Background Technology

[0002] Weight and body size are important indicators for judging the growth, development, health status, feed conversion efficiency, and slaughter time of goats. Traditional weight measurement usually requires driving goats to a weighbridge or weighing pen, while body size measurement requires manual use of tools such as measuring tapes and measuring sticks to touch the animal's body surface, which is labor-intensive and can easily cause stress.

[0003] While two-dimensional image methods can reduce contact, they lack depth information and struggle to reliably reflect three-dimensional geometric information such as chest circumference, abdominal circumference, chest width, rump width, volume, and surface area. Three-dimensional point cloud-based methods can directly describe the geometric morphology of animal bodies, but current research primarily focuses on large livestock such as cattle and pigs. Further improvements are needed for goats, which are smaller, have blurred chest and abdominal boundaries, fewer point clouds on their limbs, and localized missing point clouds.

[0004] Existing techniques for measuring the body size of animals using 3D point clouds still have the following shortcomings: First, existing methods are mostly focused on large livestock such as cattle and pigs. Goats are smaller, with blurred chest and abdominal boundaries, fewer points on their limbs, and are prone to local point cloud omissions. Directly applying these methods can easily lead to cross-sectional deviations, skewed body height, and inaccurate start and end points of body length. Second, existing point cloud segmentation methods are insufficient in representing the features of small target areas such as chest and abdominal boundaries, rump boundaries, and limb ends. They are easily affected by uneven point cloud density and class imbalance, resulting in missing limb segments and unstable boundaries. Third, existing body size extraction methods often use a single cross-section, convex hull contour, or straight-line distance between the head and tail. When the point cloud is incomplete or the backline is curved, it can easily lead to circumference breaks, width interference from external points, and underestimation of body length. Fourth, weight estimation relies heavily on a small number of body size parameters and fails to fully utilize 3D morphological information such as volume and surface area, resulting in insufficient estimation stability and generalization ability.

[0005] Therefore, there is an urgent need for a non-contact method and system for estimating the body size and weight of goats based on enhanced point cloud segmentation to overcome the problems existing in the current technology. Summary of the Invention

[0006] The purpose of this invention is to provide a non-contact method and system for estimating the body size and weight of goats based on enhanced point cloud segmentation. This addresses the problems of traditional manual measurement being labor-intensive and prone to stress, the lack of three-dimensional geometric information in two-dimensional images, the fact that existing point cloud research focuses on large livestock such as cattle and pigs and lacks a specific solution for the body shape characteristics of goats, and the tendency to directly apply existing methods to goat point clouds to cause cross-section selection bias, body height skewness, and inaccurate start and end points of body length.

[0007] To achieve the above objectives, the present invention provides a non-contact method for estimating the body size and weight of goats based on enhanced point cloud segmentation, comprising the following steps: S1, data acquisition and preprocessing: acquiring three-dimensional point cloud data of the target goat, removing outliers, normalizing coordinates and sampling the three-dimensional point cloud data to obtain a standardized point cloud; S2. Construct and train a GKA-enhanced point cloud segmentation network to obtain point-by-point region labels: Input the standardized point cloud into the trained GKA-enhanced point cloud segmentation network to obtain point-by-point region labels including head and neck, chest, abdomen, rump, left forelimb distal end, right forelimb distal end, left hindlimb distal end, and right hindlimb distal end; The GKA-enhanced point cloud segmentation network includes: a local geometry aggregation module, a group attention enhancement module, a learnable spline mapping module, and a point-by-point classification module; All point clouds with point-by-point region labels are segmented point clouds; S3. Establish body axis, width axis, height axis and candidate areas for body size measurement: Establish the body axis, width axis and height axis of the goat based on point-by-point part labels, and determine the candidate areas for body size measurement based on the chest, abdomen, rump and limb end cloud. S4. Calculate chest circumference, abdominal circumference, chest width, and hip width: Perform multi-section extraction and curve fitting within the candidate area for body size measurement, and combine coverage filtering and robust statistics to obtain chest circumference, abdominal circumference, chest width, and hip width. S5. Calculate body length and height: Obtain body length and height based on the three-dimensional B-spline curve of eight key points on the back and the horizontal support surface; S6. Calculate volume and surface area based on segmented point cloud: Calculate the total volume and local volume using one or more of voxel filling, Poisson surface reconstruction, or alpha shape reconstruction; calculate the surface area by summing the area of ​​the facets of the reconstructed triangular mesh; use geometric compensation for missing point cloud areas and retain confidence markers. S7. Construct morphological feature vectors and calculate weight estimates: Construct morphological feature vectors based on body size parameters, volume and surface area parameters, and calculate the weight estimates of the target goat using a regression model; the morphological feature vectors include at least six of the following: body length, body height, chest circumference, abdominal circumference, chest width, rump width, total volume, local volume and surface area.

[0008] S8. Confidence Assessment and Result Output: Point cloud integrity, segmentation confidence, and measurement cross-section validity are scored. Each score is normalized to 0-1 or 0-100 and weighted using preset weights. When the measurement cross-section contour coverage is greater than a set threshold and the segmentation result quality score is greater than a set threshold, the corresponding measurement result is considered reliable; otherwise, the relevant measurement result is marked as low confidence or a re-acquisition is prompted. Simultaneously, the estimated body weight, body size parameters, point cloud integrity score, segmentation confidence score, and measurement cross-section validity score are output.

[0009] Preferably, in S2, the GKA-enhanced point cloud segmentation network uses a PointNeXt-type local aggregation encoder as its backbone, sets a group attention enhancement module after the local geometric aggregation module, and sets a learnable spline mapping module in the feature propagation stage or the point-by-point classification stage; the group attention enhancement module performs group statistics, normalization, weight generation, and residual recalibration on the point cloud features after local geometric aggregation to enhance the feature representation of small target points at the chest and abdomen boundaries and limbs; the group attention enhancement module satisfy Where F is the input feature, This indicates grouped pooling statistics. Represents a lightweight mapping function. This represents the Sigmoid function. This can be either a learnable residual coefficient or a preset scaling factor.

[0010] Preferably, the learnable spline mapping module in S2 uses B-spline basis functions and learnable coefficients to form a nonlinear feature transformation. This nonlinear feature transformation can be KAN spline activation, a UnitKAN feature propagation layer, or a channel-wise mapping formed by combining linear and spline terms. The learnable spline mapping... satisfy ,in, For B-spline basis functions, Here are the learnable spline coefficients, and w is the linear weight. As input features, This is the index of the B-spline basis function.

[0011] Preferably, obtaining the chest and abdominal circumference in S4 includes the following steps: determining a central section within the candidate chest or abdominal region, extracting multiple neighboring sections in the front-back direction of the central section; performing two-dimensional projection and closed B-spline curve fitting on the point cloud of each neighboring section; removing sections with point cloud coverage below a threshold; and taking the median of the curve perimeter of the remaining sections as the chest or abdominal circumference.

[0012] Preferably, obtaining the chest width and rump width in S4 includes the following steps: extracting corresponding candidate point clouds based on point-by-point labels of the chest and rump, calculating the quantile bandwidth along the width axis in the corresponding cross section or local point cloud, and using it as the chest width or rump width; the chest width and rump width represent the forequarter width and hindquarter width, respectively, which together reflect the lateral development status of the goat's trunk, with the rump width participating in weight estimation as a hindquarter width phenotypic parameter.

[0013] Preferably, the central section is determined by one or more of the following: a linear support vector machine, a part boundary projection, or a body axis geometric constraint. The linear support vector machine divides the front and back samples by the projection position of the candidate point in the body axis direction and outputs the normal vector of the candidate cutting surface.

[0014] Preferably, obtaining the body length includes the following steps: extracting eight key points distributed along the back direction from the posterior edge of the head and neck, the anterior part of the chest, the middle part of the chest, the thoracoabdominal transition zone, the middle part of the abdomen, the posterior part of the abdomen, the anterior part of the buttocks, and the posterior edge of the buttocks; fitting a three-dimensional B-spline curve to the eight key points; and calculating the body length based on the arc length of the three-dimensional B-spline curve.

[0015] Preferably, the acquisition of body height includes the following steps: determining the horizontal support surface based on the lowest quantile of the limb endpoint cloud, determining the measurement key point from the highest point of the chest or scapular region, and calculating the vertical distance from the measurement key point to the horizontal support surface as the body height.

[0016] Preferably, the training and evaluation of the GKA-enhanced point cloud segmentation network employ one or more of the following metrics: Focal Loss, mean Intersection over Union (mIoU), overall accuracy (OA), and mean class accuracy (mAcc). , , ;in, Let c be the number of points that were correctly segmented in class c. The number of points that were incorrectly classified into class c. This represents the number of points in class c that were not correctly identified. The total number of categories, This represents the total number of points in the entire point cloud.

[0017] This invention also provides a non-contact goat body size and weight estimation system based on enhanced point cloud segmentation, comprising: a point cloud acquisition module, a preprocessing module, a GKA enhanced point cloud segmentation module, a body size measurement module, a volume and surface area calculation module, a weight estimation module, a quality control module, and a result output module; wherein: The point cloud acquisition module is connected to the preprocessing module to acquire the three-dimensional point cloud data of the target goat; The preprocessing module is connected to the point cloud acquisition module and the GKA enhanced point cloud segmentation module. It is used to remove outliers, normalize coordinates, and sample 3D point cloud data to obtain a standardized point cloud. The GKA enhanced point cloud segmentation module is connected to the preprocessing module, body size measurement module, and volume and surface area calculation module. It is used to output point-by-point part labels for the head and neck, chest, abdomen, buttocks, left forelimb end, right forelimb end, left hindlimb end, and right hindlimb end to obtain segmented point clouds. The body size measurement module is connected to the GKA enhanced point cloud segmentation module and the weight estimation module to obtain chest circumference, abdominal circumference, chest width, hip width, body length and body height; The volume and surface area calculation module is connected to the GKA enhanced point cloud segmentation module and the weight estimation module to calculate the total volume, local volume and surface area. The weight estimation module is connected to the body size measurement module, volume and surface area calculation module, quality control module and result output module, and is used to construct morphological feature vectors and regress to output weight estimation values; The quality control module is connected to the preprocessing module, the GKA-enhanced point cloud segmentation module, the body size measurement module, the volume and surface area calculation module, the weight estimation module, and the result output module. It is used to evaluate the integrity of the point cloud, the segmentation quality, and the rationality of the body size, and output scores and low confidence prompts. The results output module is connected to the weight estimation module and the quality control module to synchronously output weight estimates, body size parameters, various scores, and visualization results.

[0018] Therefore, the present invention employs the above-mentioned non-contact goat body size and weight estimation method and system based on enhanced point cloud segmentation, which has the following beneficial effects: (1) This invention extracts chest and abdominal circumference by multi-section B-spline curve fitting and median statistics. Multiple adjacent sections are extracted in the candidate area of ​​chest or abdomen and closed B-spline fitting is performed. Combined with coverage screening and median statistics, it effectively solves the problem of error amplification caused by slight offset of section position in single section measurement, and the problem of circumference breakage or excessive expansion caused by local point cloud gaps and outliers. It significantly improves the stability and repeatability of circumference measurement under residual point cloud conditions.

[0019] (2) The present invention obtains body length by fitting a three-dimensional B-spline curve to eight key back points from the posterior edge of the head and neck to the posterior edge of the rump and calculating the arc length. Unlike the traditional straight-line distance measurement, it can truly reflect the natural curvature of the goat's back and solves the problem of insufficient expression of body development characteristics caused by ignoring the curvature of the back line in the straight body length. This makes the body length parameters more consistent with the real trunk shape of the goat.

[0020] (3) Based on the point-by-point label of the hindquarters, the present invention extracts the candidate point cloud of the hindquarters, calculates the quantile bandwidth along the width axis as the hindquarters width, and incorporates the hindquarters width as an independent phenotypic parameter into the weight estimation, which solves the problem that the existing methods only focus on body length, body height and chest circumference and ignore the development status of the hindquarters of goats, making the weight estimation feature system more complete.

[0021] (4) The present invention constructs a horizontal support surface by using the lowest quantile point in the limb region and calculates the vertical distance from the highest point in the chest or scapular region to obtain the body height. This is different from the traditional oblique plane fitting method and solves the problems of insufficient number of points at the limb ends, ground noise and occlusion causing body height skewness, thus improving the reliability of body height measurement.

[0022] (5) This invention combines at least six parameters from body length, body height, chest circumference, abdominal circumference, chest width, rump width, total volume, local volume and surface area into a three-dimensional morphological feature vector, and outputs the weight estimate through a regression model. It makes full use of the spatial geometric information contained in the three-dimensional point cloud, solves the problem that existing weight estimation relies on a single or a small number of body size features, resulting in insufficient generalization ability, and can adapt to the measurement needs of goats of different breeds, different growth stages and different body shapes.

[0023] (6) This invention forms a complete three-dimensional phenotypic measurement chain from point cloud acquisition, part segmentation, body size measurement, volume and surface area calculation to weight estimation. The entire process does not require manual contact or fixing of goats, which solves the problems of high labor intensity, low efficiency and easy to cause animal stress in traditional manual measurement. It is suitable for individual growth monitoring, group management and breeding assistance in large-scale farms.

[0024] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0025] Figure 1 This is a schematic diagram of the depth camera acquisition device for the non-contact goat body size and weight estimation system based on enhanced point cloud segmentation of the present invention. Figure 2 This is a schematic diagram of the overall technical roadmap for the non-contact body size and weight estimation method for goats based on enhanced point cloud segmentation, as described in this invention. Figure 3 This is a schematic diagram of the GKA-enhanced point cloud segmentation network structure for the non-contact goat body size and weight estimation method based on enhanced point cloud segmentation of the present invention. Figure 4 This is a flowchart of the body size measurement method for the non-contact body size and weight estimation method for goats based on enhanced point cloud segmentation, as described in this invention. Figure 5 This is a flowchart of the volume and surface area calculation steps for the non-contact goat body size and weight estimation method based on enhanced point cloud segmentation of the present invention. Figure 6 This is a schematic diagram of the structure of the non-contact goat body size and weight estimation system based on enhanced point cloud segmentation, which is the application of the present invention. Figure 7 This is a visualization of the chest circumference section curve fitting for the non-contact goat body size and weight estimation method based on enhanced point cloud segmentation of the present invention. Figure 8 This is a visualization of the abdominal circumference section curve fitting of the non-contact goat body size and weight estimation method based on enhanced point cloud segmentation of the present invention. Figure 9 The images show visualizations of body height and body length measurements for the non-contact goat body size and weight estimation method based on enhanced point cloud segmentation according to the present invention. (a) is a visualization of body height measurement, and (b) is a visualization of body length measurement. Detailed Implementation

[0026] The following detailed description of embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0027] Example 1 Please see Figures 1-9 A non-contact method for estimating the body size and weight of goats based on enhanced point cloud segmentation includes the following steps: S1. Data Acquisition and Preprocessing: Acquire the 3D point cloud data of the target goat, and perform outlier removal, coordinate normalization, and sampling on the 3D point cloud data to obtain a standardized point cloud.

[0028] Specifically, the preprocessing module performs field validation on the input point cloud, removing points with missing coordinates, missing labels, non-numerical coordinate values, or coordinates outside the acquisition space. For duplicate points formed by multi-view fusion, voxel grid downsampling or radius neighborhood deduplication is used to reduce redundancy, making the point cloud density more balanced in the same surface region.

[0029] To address the orientation differences of goats in the collection channel, the body axis is estimated based on the principal component direction of the trunk point cloud, and the direction from the head and neck to the rump is aligned with the unified body axis direction. When the head and neck point cloud is missing, the direction is determined by the relative position of the centroids of the chest, abdomen, and rump along the principal axis, in order to reduce the errors caused by the confusion of coordinate directions in the subsequent body length starting point, chest and abdomen section, and rump section.

[0030] The normalization process involves translating the point cloud to near the centroid and scaling it according to the maximum radius, torso length, or device calibration scale. Normalized coordinates are used as network input during model training and then mapped back to the true centimeter scale during body size measurement to balance network training stability with the physical meaning of the actual body size output. For data augmentation, small rotations around the height axis, coordinate perturbations, and local noise disturbances are used, avoiding strong scaling, mirroring, or large-area random point loss that would alter the true body shape. For categories with fewer limb endpoints, class weights are introduced into the loss function instead of altering the original geometric distribution by replicating limb endpoints.

[0031] S2. Construct and train a GKA-enhanced point cloud segmentation network to obtain point-by-point region labels: Input the standardized point cloud into the trained GKA-enhanced point cloud segmentation network. Obtain point-by-point region labels including head and neck, chest, abdomen, rump, left forelimb distal end, right forelimb distal end, left hindlimb distal end, and right hindlimb distal end; the GKA-enhanced point cloud segmentation network includes: a local geometry aggregation module, a group attention enhancement module, a learnable spline mapping module, and a point-by-point classification module; the entire point cloud with point-by-point region labels is the segmented point cloud.

[0032] The GKA-enhanced point cloud segmentation network uses a PointNeXt-type local aggregation encoder as its backbone. A group attention enhancement module is added after the local geometric aggregation module, and a learnable spline mapping module is added during the feature propagation or point-by-point classification stage. Specifically, the GKA-enhanced point cloud segmentation network is denoted as GKA-NeXt. It uses a PointNeXt-type local aggregation encoder as its backbone, introduces a group attention enhancement module after local neighborhood aggregation at various levels in the local geometric aggregation module, and introduces KAN spline mapping in the feature propagation or point-by-point classification stage, forming the GKA-NeXt+KAN model.

[0033] The grouped attention enhancement module performs grouped statistics, normalization, weight generation, and residual recalibration on the point cloud features after local geometric aggregation to enhance the feature representation of small target points at the chest and abdomen boundaries and limbs. satisfy Where F is the input feature, This indicates grouped pooling statistics. Represents a lightweight mapping function. This represents the Sigmoid function. The residual coefficients can be learned or preset scaling factors. This structure can enhance difficult-to-distinguish regions such as the chest and abdomen boundaries, buttock boundaries, and extremities without significantly increasing the number of parameters. A grouped attention enhancement module is inserted after local feature aggregation. This module first divides the channel features into several groups, then calculates statistics such as the mean, maximum, or standard deviation within each group. Attention weights are obtained through lightweight mapping and applied to the original features using residual recalibration. Because the residual branches have relatively small initial weights, existing geometric features are not destroyed during the initial training phase, and difficult-to-distinguish regions are automatically enhanced after training.

[0034] The learnable spline mapping module uses B-spline basis functions and learnable coefficients to form a nonlinear feature transformation. This nonlinear feature transformation can be KAN spline activation, a UnitKAN feature propagation layer, or a channel-wise mapping formed by combining linear and spline terms. satisfy ,in, For B-spline basis functions, Here are the learnable spline coefficients, and w is the linear weight. As input features, This layer represents the index of the B-spline basis functions. This layer is an optional implementation, replacing the traditional fixed nonlinear activation with a family of functions formed by combining B-spline basis functions and learnable coefficients, used in the feature propagation stage, the classification head stage, or a combination of both. Specific implementations of the learnable spline mapping module include: setting a UnitKAN layer in the feature propagation layer and the point-by-point classification head, without replacing the entire multilayer perceptron; using third-order B-splines with 8 to 12 control points; introducing L2 regularization terms to the spline coefficients to reduce the risk of overfitting under small sample conditions; training the PointNeXt / GKA backbone network first, then fine-tuning the KAN spline layers with a smaller learning rate. Regarding the loss function setting, Focal Loss and class-adaptive weights can be used together to increase the weights for difficult-to-classify regions such as the extremities, the chest and abdomen transition zone, and the buttock boundary; and neighborhood consistency smoothing is applied to low-confidence points during the inference stage without changing the original point cloud coordinates. The above implementation retains the ability of GKA attention enhancement to express local boundaries, while also utilizing KAN spline mapping to improve the nonlinear fitting ability of complex surface features.

[0035] For lightweight deployment, the GKA-enhanced point cloud segmentation network can be constructed as a lightweight point cloud segmentation network. It employs a PointNeXt-like encoder to extract local geometric features and includes a multi-level sampling aggregation structure: the first stage retains a high-resolution local point set for learning chest and abdominal boundaries, leg ends, and head and neck details; subsequent stages gradually expand the neighborhood radius and receptive field to learn the overall morphology of the torso and its anterior-posterior orientation; the decoding stage restores deep semantic features to point-by-point resolution through feature propagation. After local geometric aggregation, the network incorporates an attention enhancement module to group and statistically analyze point cloud features, generating weights by combining average and extreme value information. Residual methods are used to recalibrate the original features, ensuring stability in the initial stage. After training, this enhances the chest and abdominal boundaries, buttock boundaries, and small target regions at the extremities.

[0036] The GKA-enhanced point cloud segmentation network is trained using Focal Loss and class weights, with the loss function... for: ; in, The probability is predicted for the true class, and γ is the focusing factor. This represents the category weights. For categories with fewer points, such as limbs and extremities, The weights are determined by the ratio of the median class frequency to the current class frequency to reduce the underclassing of small targets caused by class imbalance. When training the segmentation network, cross-entropy loss, Focal Loss, Dice Loss, or a weighted combination thereof are used. For small targets such as chest and abdominal boundaries and limbs, adaptive weights are set based on the proportion of points in each class to ensure that the minority class obtains sufficient gradients during backpropagation. After training, the model outputs the probability of each point belonging to one of the eight body parts, and the class with the highest probability is used as the final body part label.

[0037] In terms of model evaluation, the GKA-enhanced point cloud segmentation network is evaluated using one or more methods among overall accuracy, mean intersection-over-union ratio (CIU), CIU for each category, and confusion matrix; body size measurement is evaluated using one or more methods among mean absolute error (MAE), mean absolute percentage error (MAS), and correlation coefficient; and weight estimation is evaluated using one or more methods among mean absolute error (MAE), root mean square error (RMSE), and coefficient of determination (CDE). These metrics are used to evaluate system performance and do not limit the scope of protection of this invention.

[0038] When used in experimental descriptions for embodiments of the invention, mIoU, OA, mAcc, and Acc are used to evaluate the GKA-enhanced point cloud segmentation network; mean absolute error (MAE), root mean square error (RMSE), mean absolute percentage error (MAPE), and coefficient of determination (R²) are used. 2 Used to evaluate body size or weight estimation. Segmentation performance is evaluated using the mean intersection-union ratio (mIoU), overall accuracy (OA) after combining all valid points, the arithmetic mean (Acc) of point-by-point accuracy for each goat sample, the mean class precision (mAcc), and the class intersection-union ratio. ,in: ; ; ; ; ; in, Let c be the number of points that were correctly segmented in class c. The number of points that were incorrectly classified into class c. This represents the number of points in class c that were not correctly identified. The total number of categories, This represents the total number of points in the entire point cloud. The total number of test samples, Let be the number of points that are correctly segmented in the i-th goat. Let be the total number of points in the point cloud for the i-th goat.

[0039] The calculation methods for each evaluation index in body size measurement and weight estimation are as follows: ; ; ; ; in, Let i be the true value of the i-th sample. Let i be the estimated value of the i-th sample. is the mean of the true values, and M is the total number of samples.

[0040] The eight body part labels of this invention can also be expanded or combined according to the needs of the farm. For example, the head and neck can be further divided into head and neck, the torso into shoulders, back, abdomen and rump, or the extremities can be combined into a leg support area. Changes in body part categories do not affect the core idea of ​​establishing candidate measurement regions based on label point clouds.

[0041] To facilitate deployment at the edge of aquaculture farms, this invention does not limit the network to a large model. The network can be configured with different channel widths and layers based on the device's computing power, and inference latency can be reduced through pruning, quantization, or distillation. Any alternative implementation of this invention is possible as long as it can output point-by-point labels for the head and neck, chest, abdomen, rump, and extremities that meet measurement requirements.

[0042] S3. Establish candidate regions for body axis, width axis, height axis, and body size measurement: Based on point-by-point site labels, establish the goat's body axis, width axis, and height axis, and determine candidate regions for body size measurement based on the cloud data of the chest, abdomen, rump, and limb ends. For areas where the chest and abdomen boundaries overlap or have low segmentation confidence, a joint screening is performed using the quantile position along the body axis direction, the centroid of adjacent sites, and the number of local cross-sectional points to avoid cross-sections falling in locations with excessively large gaps or severe label mixing.

[0043] S4. Calculate chest circumference, abdominal circumference, chest width, and hip width: Perform multi-section extraction and curve fitting within the candidate area for body size measurement, and combine coverage filtering and robust statistics to obtain chest circumference, abdominal circumference, chest width, and hip width.

[0044] The acquisition of chest and abdominal circumference in S4 includes the following steps: A central section is determined within the candidate chest or abdominal region, and multiple adjacent sections are extracted along the body axis in the front-to-back direction of the central section. Each thin section has the same thickness, or the thickness is adaptively adjusted based on the local point cloud density. For sections with fewer points than a threshold, insufficient closed coverage angle, or severe missing points on the left and right sides, the measurement module marks them as invalid sections and excludes them from the final circumference statistics. Within each valid section, the measurement module projects the 3D section points onto a 2D local coordinate system perpendicular to the section normal vector and fits the projected points with a closed B-spline curve. During the fitting process, outliers are filtered by radius or quantile truncation, the curve order is selected as third order, and the number of control points is adaptively determined based on the number of section points and curvature changes. The circumference calculation is not simply connecting the outer contour of the point cloud, but rather uniformly sampling along the fitted closed curve and calculating the sum of distances between adjacent sampling points. This reduces the problem of broken lines caused by local gaps in the point cloud and avoids excessive expansion in the abdominal or chest concave areas caused by the convex hull method. Cross sections with point cloud coverage below the threshold are removed. For the remaining cross sections, the final chest or abdominal circumference is calculated using the median of the perimeters of multiple effective cross sections, or the truncated mean or confidence-weighted mean.

[0045] The closed curves of the chest and abdominal circumference sections can be represented as follows: ,in For the control points of the cross-sectional curve, These are third-order B-spline basis functions. The gravitational dimension can be expressed as... In actual calculations, samples can be uniformly taken along the closed curve, and the distances between adjacent sampling points can be accumulated. For multiple effective sections, the final girth can be expressed as... ,in, For the first The girth value of each effective cross section, =1,2,…, K , K This represents the total number of effective cross sections.

[0046] like Figure 7 and Figure 8 As shown, after projecting the three-dimensional point cloud of the candidate cross-sections of the chest and abdomen onto the local two-dimensional coordinate system, the cross-section point cloud is distributed along the outer contour of the goat's torso. After outlier processing and B-spline curve fitting, the fitted curve can more smoothly characterize the cross-sectional shape and reduce the influence of missing local point cloud and discrete noise on the girth calculation, thus providing a stable geometric contour for subsequent calculation of chest and abdominal circumference along the fitted curve.

[0047] The acquisition of chest width and rump width in S4 involves the following steps: Candidate point clouds are extracted based on point-by-point labels of the chest and rump. The quantile bandwidth is calculated along the width axis in the corresponding cross-section or local point cloud and used as the chest width or rump width, reducing the influence of missing lateral points and outliers. Chest width and rump width represent the forequarter width and hindquarter width, respectively, jointly reflecting the lateral development of the goat's trunk. Rump width serves as a phenotypic parameter for hindquarter width in weight estimation. A representative width between the 30th and 70th percentiles in the width direction is taken, or the 20th to 80th percentile range is selected based on the completeness of the point cloud. This method reduces the influence of unilateral lateral points, leg remnants, and boundary defects on width measurement.

[0048] Among them, chest width and hip width can be expressed as: Where Y is the projected coordinate of the local cross-sectional point cloud on the width axis. and The preset quantiles are used, for example, 70% and 30%. This quantile bandwidth formula can reduce the impact of outliers and one-sided missing values ​​on the width estimation.

[0049] The central section is determined by one or more of the following: linear support vector machine, part boundary projection, or body axis geometric constraints. The linear support vector machine divides the front and back samples by the projection position of the candidate point in the body axis direction and outputs the normal vector of the candidate cutting surface.

[0050] S5. Calculate body length and height: Obtain body length and height based on the three-dimensional B-spline curve of eight key points on the back and the horizontal support surface.

[0051] The body length acquisition process includes the following steps: Eight key points are extracted from the posterior edge of the head and neck, the anterior chest, the middle chest, the thorax-abdomen transition zone, the middle abdomen, the posterior abdomen, the anterior rump, and the posterior edge of the rump, distributed along the back direction. The height of the key points is selected from the upper quantile points of the corresponding local areas rather than the highest single point to avoid curve elevation caused by hair tips or outliers. The eight key points are sorted according to the body axis direction and fitted with a three-dimensional B-spline curve. The curve arc length is calculated to obtain the body length. This method differs from traditional straight-line distance measurement and can accurately reflect the natural curvature of the goat's back. For applications requiring only the straight body length, the projected distance of the front and rear key points along the body axis direction is directly calculated. This invention simultaneously covers both curved body length and straight body length output formats.

[0052] The length curve of the three-dimensional B-spline volume can be represented as follows: ,in These are key points or control points on the back. The basis functions are third-order B-spline functions. The body length BL can be integraled from the arc length. It can also be obtained by uniform sampling. ,in, The parameter value corresponding to the j-th sampling point. Let be the position vector of the j-th sampling point on the 3D B-spline curve.

[0053] Obtaining body height involves the following steps: determining the horizontal support surface based on the lowest quantile of the limb endpoint cloud, which is the point cloud of the four limb ends. Since the collected point cloud may contain missing leg areas and local noise, this invention does not directly use all the lowest points of the legs. Instead, stable low quantiles are taken in the four limb end regions, and then the horizontal support surface is constructed based on these points. If the device coordinate system has already been calibrated to determine the true direction of gravity, the normal vector of the support surface is directly taken as the height axis direction. Key measurement points for the shoulder or cross-section are determined from the upper edge of the chest, near the scapula, or in the neighborhood of key points in the forearm. The highest point in the chest or scapula region is used to determine the measurement key points. The upper quantile of height or the smoothed highest point is taken within the local area, and the vertical distance from the measurement key point to the horizontal support surface is calculated as the body height. This avoids skewed body height direction due to oblique plane fitting, thus ensuring that the body height direction is consistent with the foot support surface.

[0054] like Figure 9 As shown, eight key points extracted along the back of the goat are fitted with three-dimensional B-splines to form a body length curve that varies with the natural undulation of the back. Compared with directly calculating the straight distance between the front and rear endpoints, this curve can more fully reflect the curvature of the back. At the same time, by determining the horizontal support surface using the stable low quantile point at the limb end and calculating the vertical distance from the key point in the chest or scapular region to the support surface, the body height consistent with the foot support direction can be obtained.

[0055] S6. Volume and Surface Area Calculated Based on Segmented Point Cloud: Volume is calculated using a fast voxel filling path and a high-precision surface reconstruction path. The fast voxel filling path establishes bounding boxes and voxels for the segmented point cloud, and determines the internal voxels through multi-directional rays or neighborhood connectivity; the high-precision surface reconstruction path first estimates the point cloud normal vectors, and then performs Poisson surface reconstruction or alpha shape reconstruction to obtain a closed or approximately closed mesh.

[0056] Surface area calculation is based on the sum of the areas of the reconstructed triangular mesh, or the surface area is calculated separately for each local area and then combined. For areas with severe point cloud defects, geometric compensation is performed using the opposite side of the same goat or adjacent sections, but the compensation results should retain confidence markers so that the weight estimation module can distinguish between true observation features and compensated features.

[0057] S7. Constructing Morphological Feature Vectors and Calculating Weight Estimates: Based on body size parameters, volume, and surface area parameters, construct morphological feature vectors and use a regression model to calculate the estimated weight of the target goat. The morphological feature vectors include at least six of the following: body length, body height, chest circumference, abdominal circumference, chest width, rump width, total volume, local volume, and surface area. Outlier truncation, standardization, and correlation screening are performed before feature input. The regression model uses linear regression, ridge regression, random forest, gradient boosting tree, support vector regression, or lightweight multilayer perceptron. For datasets with small sample sizes, cross-validation and feature importance analysis are used to reduce the risk of overfitting.

[0058] S8. Confidence Assessment and Result Output: Point cloud integrity, segmentation confidence, and measurement cross-section validity are scored. Specifically: The point cloud integrity score is determined based on the number of valid points, the spatial coverage of the torso surface, the presence of large-area voids, and the integrity of the support areas at the limb ends; the segmentation confidence score is determined based on the maximum class probability per point output by the network, the proportion of points in each part, the continuity of adjacent part boundaries, and the class confidence of the limb ends; the measurement cross-section validity score is determined based on the number of valid points in the cross-section, the cross-section contour coverage, the integrity of the left and right sides, and the fitting residuals. All scores are normalized to 0-1 or 0-100 and weighted using preset weights; when the measurement cross-section contour coverage... And the segmentation result quality score A score of 0.8 indicates the measurement result is reliable; otherwise, the measurement result is marked as low confidence or a re-acquisition is prompted. The system simultaneously outputs the estimated body weight, body size parameters, point cloud integrity score, segmentation confidence score, and measurement cross-section validity score.

[0059] For the segmentation results, the system calculates the percentage of points in each body part, the continuity of boundaries between adjacent body parts, and the confidence level of small targets at the extremities. If the percentage of points in the chest, abdomen, or buttocks deviates significantly from the distribution of samples in the same batch, the system will indicate an abnormal segmentation. If there are insufficient points at the extremities, the height measurement module can reduce the confidence level of the corresponding results or require re-collection.

[0060] For body size results, the system sets biologically reasonable range constraints and temporal continuity constraints. For example, the body length, body height, chest width, and rump width of the same goat on adjacent sampling days should not show abrupt changes that do not conform to the growth pattern; chest circumference and abdominal circumference may fluctuate to some extent with feeding status, but should still remain basically consistent with body shape, weight, and volume characteristics.

[0061] To facilitate manual review, the system saves visualizations of chest circumference, abdominal circumference, body width, back length curves, and body height support surfaces, as well as key attached diagrams such as width quantile bandwidth. Reviewers can use these key sections and curves to determine whether the measurement location falls within a reasonable area, eliminating the need to check the entire 3D point cloud point by point, thus reducing the workload of manual review. The visualization results are only used to interpret the measurement process and assist in manual review; they are not involved in the final body size calculation and do not alter the robust statistics and curve fitting process described above.

[0062] Example 2 This embodiment provides a non-contact goat body size and weight estimation system based on enhanced point cloud segmentation, including the following modules: Point cloud acquisition module: Includes a depth camera or multi-view depth acquisition device, used to acquire 3D point cloud data of the target goat. The acquisition device is fixedly installed near the goat's passage or measurement enclosure. The point cloud acquisition device of this invention includes, but is not limited to, at least one of a structured light camera, a binocular camera, a ToF camera, or a lidar; the acquisition method of the point cloud acquisition device includes any one of single-view acquisition, multi-view acquisition, or moving scan acquisition, used to acquire 3D point cloud data of the target goat's body surface.

[0063] The preprocessing module removes points containing NaN, with abnormal column counts, or obvious outliers. It normalizes the data based on the centroid and maximum radius, and uses farthest-point sampling or voxel sampling to obtain a standardized point cloud with a uniform number of points. During the training phase, random rotation around the height axis and small-amplitude Gaussian coordinate perturbations are used for data augmentation to avoid strong flipping, overscaling, or random point loss that could disrupt the true size distribution of the goats.

[0064] GKA Enhanced Point Cloud Part Segmentation Module: This module is used to input standardized point clouds into the GKA Enhanced Point Cloud Part Segmentation Network to obtain point-by-point part labels for the head and neck, chest, abdomen, rump, left forelimb tip, right forelimb tip, left hindlimb tip, and right hindlimb tip.

[0065] Body size measurement module: It is used to establish the body axis, width axis and height axis of goat based on point-by-point part labels, determine the candidate area for body size measurement, and perform multi-section extraction and curve fitting within the candidate area. Combined with coverage screening and robust statistics, it obtains chest circumference, abdominal circumference, chest width and buttock width; it is also used to obtain body length and body height based on the three-dimensional B-spline curve of eight key points on the back and the horizontal support surface of the limb ends.

[0066] Volume and Surface Area Calculation Module: Used to calculate the total and local volumes based on one or more of the following methods: voxel filling, Poisson surface reconstruction, or alpha shape reconstruction, based on segmented point clouds; and to calculate the surface area based on the sum of the facet areas of the reconstructed triangular mesh.

[0067] Weight estimation module: It is used to combine at least six of the following into a morphological feature vector: body length, body height, chest circumference, abdominal circumference, chest width, rump width, total volume, local volume, and body surface area, and output the weight estimation value of the target goat using any one of the following regression models: linear regression, random forest, gradient boosting tree, support vector regression, or lightweight multilayer perceptron.

[0068] The quality control module is used to check the integrity of the point cloud, evaluate the quality of the segmentation results, and verify the reasonableness of the volumetric dimensions. The checks include whether the number of points is below a threshold, whether there are large areas of voids, whether the chest and abdominal labels are significantly misaligned, whether the extremities can form support areas, and whether the coordinate axis directions are consistent with expectations. If any check item fails to meet the preset conditions, the system can mark the sample as pending review.

[0069] The results output module synchronously outputs estimated weight, body size parameters, point cloud integrity score, segmentation confidence score, and measurement cross-section validity score. When there are many missing point cloud data, insufficient limb support points, or insufficient chest and abdominal cross-section coverage, the results are marked as low confidence or a re-acquisition is prompted. This module displays the segmentation results, body size annotations, volume distribution, estimated weight, and historical trends on a local industrial control computer, embedded GPU, or host computer software. It also saves the visualization results of chest circumference, abdominal circumference, body width, back length curve, and body height support surface, facilitating manual verification by using key cross-sections and key curves to determine if the measurement location falls within a reasonable area, eliminating the need to check the entire 3D point cloud point by point. Simultaneously, the module establishes an index using individual ID, acquisition date, and device ID to ensure correct correspondence between point cloud, body size, and weight information for the same goat at different time points. Each goat corresponds to an original point cloud file, a cleaned point cloud file, a point-by-point label file, a manual body size record, a weighing record, and a measurement visualization file. Simultaneously, the module connects with electronic ear tags, a pen management system, and a growth record database to achieve long-term tracking at the individual level. Based on estimated weight and body size changes, it generates growth curves, abnormal warnings, and grouping suggestions for use in large-scale farms for growth monitoring, grouping management, and breeding assistance.

[0070] This invention can be deployed in fixed measuring pens, mobile measuring vehicles, or farm walkways. Fixed measuring pens are suitable for batch weighing and periodic remeasurement of body size; mobile measuring vehicles are suitable for environmental inspections of scattered pens; and farm walkways are suitable for daily, pass-by data collection. The system can connect to electronic ear tags, pen management systems, and growth record databases to achieve long-term tracking at the individual level.

[0071] In terms of engineering operation, the system can adopt either batch processing mode or real-time mode. Batch processing mode is suitable for offline import of historical point clouds and unified calculation of body size and weight; real-time mode is suitable for automatically triggering acquisition, segmentation, measurement, and result uploading when goats pass through the acquisition channel. Both modes can use the same model and measurement module, with differences only in task scheduling and data caching methods.

[0072] For young sheep, adult sheep, different breeds of goats, or goats of different body types, adaptation can be achieved by recalibrating scale parameters, updating the regression model, or fine-tuning the segmentation network. When updating the model, existing artificial body measurements and weighing records can be used as supervised data, and newly added point clouds can be used as semi-supervised or weakly supervised data to gradually improve the generalization ability in different scenarios.

[0073] To ensure long-term stability, the system can be calibrated periodically using manually measured samples. Calibration includes adjusting depth camera extrinsic parameters, centimeter-scale conversion, body axis orientation, body size measurement thresholds, and weight regression model parameters. The system can be recalibrated when the farm changes its data acquisition equipment, changes the channel width, or changes its primary goat breed.

[0074] This invention can also be integrated with aquaculture management software to generate growth curves, abnormal warnings, and grouping suggestions based on estimated body weight and body size trends. For example, for individuals with slow body size growth but abnormally large abdominal circumference, the system can prompt further manual examination; for individuals with stable weight gain and balanced body shape, the system can record them as having normal growth status.

[0075] Example 3 The training, validation, and test sets are divided according to individual goats. Data augmentation is applied only to the training set to avoid the original and augmented point clouds of the same goat being placed in different datasets. To illustrate the model evaluation method of this invention, in a simulated evaluation embodiment, eight types of point cloud segmentation tasks are constructed using cleaned goat point clouds and their augmented samples. The eight types of point cloud segmentation tasks are trained and validated using 59 cleaned goat point clouds and their augmented samples. PointNet, PointNet++MSG (Point Network + Multi-Scale Grouping Model), PointNeXt-S (Next-Generation Point Network - Miniature Version), G-KANext (Geometric-Kolmogorov-Arnold Network Extension Model), GKA-NeXt (Grouped Attention Augmented Next-Generation Point Network), and GKA-NeXt+KAN (Grouped Attention Augmented Next-Generation Point Network + KAN Model) are selected for comparison.

[0076] The relevant comparison results are shown in Table 1, which presents a comparison of the metrics of different point cloud segmentation models in the simulation evaluation embodiment of the present invention. Each model uses the same cleaned goat point cloud and its enhanced samples to construct eight types of point cloud segmentation tasks. The evaluation metrics include overall accuracy OA, average class accuracy mAcc, average intersection-union ratio mIoU, and point-by-point accuracy Acc; among which mIoU is used as the main comparison metric.

[0077] Table 1: Comparison of metrics for different point cloud segmentation models

[0078] As shown in Table 1, under the same task settings, the GKA-NeXt+KAN model has an mIoU of 0.868, an OA of 92.3%, and an Acc of 91.9%. Compared to the PointNet++ MSG model, its mIoU is improved by 0.042, OA by 3.3 percentage points, and Acc by 3.2 percentage points; compared to the GKA-NeXt model, its mIoU is improved by 0.014, OA by 1.1 percentage points, and Acc by 1.0 percentage point. The above results are used to illustrate the index organization method of this invention; in formal implementation, they should be replaced with actual training and testing results.

[0079] Therefore, this invention employs the aforementioned non-contact goat body size and weight estimation method and system based on enhanced point cloud segmentation. It acquires goat body surface data through multi-source 3D point cloud acquisition and preprocessing, and uses a GKA enhanced point cloud segmentation network to classify the head and neck, chest, abdomen, rump, and limb extremities point by point. Based on this, body axis, width axis, and height axis are established, and candidate regions for body size measurement are determined. Chest circumference and abdominal circumference are obtained through multi-section B-spline curve fitting and coverage screening. Chest width and rump width are obtained by calculating the quantile bandwidth along the width axis. Body length is obtained by fitting 3D B-spline curves to eight key points on the back and calculating the arc length. Body height is obtained by constructing a horizontal support surface from the lowest quantile point of the limbs. Volume and surface area are calculated by combining voxel filling or surface reconstruction. Finally, at least six parameters from body length, body height, chest circumference, abdominal circumference, chest width, rump width, volume, and surface area are combined into a morphological feature vector, and a regression model outputs the estimated weight value. This invention eliminates the need for manual contact and restraint of goats, reducing the labor intensity and animal stress associated with traditional measurement methods. By employing multi-section curve fitting and median statistics, it effectively addresses the issue of single-section measurements being susceptible to cross-sectional positional shifts. The calculation of the arc length of the three-dimensional B-spline on the back accurately reflects the natural curvature of the goat's back. The construction of a horizontal support surface at the limb ends avoids skewness in the height direction, significantly improving the measurement stability and repeatability of key body size parameters such as chest circumference, abdominal circumference, body length, and body height under conditions of defects. Furthermore, by incorporating rump width, volume, and surface area into the weight estimation feature system, weight estimation becomes more accurate and reliable. This invention can be deployed in various scenarios, including fixed measurement pens, mobile measurement vehicles, or farm walkways, and is suitable for individual growth monitoring, group management, and breeding assistance in large-scale farms.

[0080] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A non-contact method for estimating the body size and weight of goats based on enhanced point cloud segmentation, characterized in that, Includes the following steps: S1. Data Acquisition and Preprocessing: Acquire the three-dimensional point cloud data of the target goat, and perform outlier removal, coordinate normalization and sampling on the three-dimensional point cloud data to obtain a standardized point cloud; S2. Construct and train the GKA-enhanced point cloud segmentation network to obtain point-by-point part labels: Input the standardized point cloud into the trained GKA-enhanced point cloud segmentation network to obtain point-by-point part labels including head and neck, chest, abdomen, buttocks, left forelimb end, right forelimb end, left hindlimb end, and right hindlimb end. The GKA-enhanced point cloud segmentation network includes: a local geometry aggregation module, a group attention enhancement module, a learnable spline mapping module, and a point-by-point classification module; the entire point cloud with point-by-point region labels is the segmented point cloud; S3. Establish body axis, width axis, height axis and candidate areas for body size measurement: Establish the body axis, width axis and height axis of the goat based on point-by-point part labels, and determine the candidate areas for body size measurement based on the chest, abdomen, rump and limb end cloud. S4. Calculate chest circumference, abdominal circumference, chest width, and hip width: Perform multi-section extraction and curve fitting within the candidate area for body size measurement, and combine coverage filtering and robust statistics to obtain chest circumference, abdominal circumference, chest width, and hip width. S5. Calculate body length and height: Based on the three-dimensional B-spline curves and horizontal support surfaces of eight key points on the back, obtain body length and height; among which, the eight key points on the back include the posterior edge of the head and neck, the anterior part of the chest, the middle part of the chest, the thoracoabdominal transition zone, the middle part of the abdomen, the posterior part of the abdomen, the anterior part of the buttocks, and the posterior edge of the buttocks. S6. Calculate volume and surface area based on segmented point cloud: Calculate the total volume and local volume using one or more of voxel filling, Poisson surface reconstruction, or alpha shape reconstruction; calculate the surface area by summing the area of ​​the facets of the reconstructed triangular mesh; use geometric compensation for missing point cloud areas and retain confidence markers. S7. Construct morphological feature vectors and calculate weight estimates: Construct morphological feature vectors based on body size parameters, volume and surface area parameters, and calculate the weight estimates of the target goat using a regression model; the morphological feature vectors include at least six of the following: body length, body height, chest circumference, abdominal circumference, chest width, rump width, total volume, local volume and surface area; S8. Confidence Assessment and Result Output: Check the integrity of the point cloud, assess the quality of the segmentation results, and verify the rationality of the body size results; simultaneously output the estimated weight, body size parameters, point cloud integrity score, segmentation confidence score, and measurement section validity score.

2. The method for non-contact body size and weight estimation of goats based on enhanced point cloud segmentation according to claim 1, characterized in that: In S2, the GKA-enhanced point cloud segmentation network uses a PointNeXt-type local aggregation encoder as its backbone. A group attention enhancement module is set after the local geometric aggregation module, and a learnable spline mapping module is set in the feature propagation stage or the point-by-point classification stage. The group attention enhancement module performs group statistics, normalization, weight generation and residual recalibration on the point cloud features after local geometric aggregation to enhance the feature representation of small target points of the chest and abdomen boundaries and limbs. Group Attention Enhancement Module satisfy Where F is the input feature, This indicates grouped pooling statistics. Represents a lightweight mapping function. This represents the Sigmoid function. This can be either a learnable residual coefficient or a preset scaling factor.

3. The method for non-contact body size and weight estimation of goats based on enhanced point cloud segmentation according to claim 2, characterized in that: The learnable spline mapping module in S2 uses B-spline basis functions and learnable coefficients to form a nonlinear feature transformation. This nonlinear feature transformation can be KAN spline activation, a UnitKAN feature propagation layer, or a channel-wise mapping formed by combining linear and spline terms. Learnable spline mapping... satisfy ,in, For B-spline basis functions, Here are the learnable spline coefficients, and w is the linear weight. As input features, This is the index of the B-spline basis function.

4. The method for non-contact body size and weight estimation of goats based on enhanced point cloud segmentation according to claim 3, characterized in that, The acquisition of chest and abdominal circumference in S4 includes the following steps: determining the central section within the candidate chest or abdominal region, extracting multiple neighboring sections in the front and back directions of the central section; performing two-dimensional projection and closed B-spline curve fitting on the point cloud of each neighboring section; removing sections with point cloud coverage below a threshold; and taking the median of the curve perimeter of the remaining sections as the chest or abdominal circumference.

5. The method for non-contact body size and weight estimation of goats based on enhanced point cloud segmentation according to claim 4, characterized in that, The acquisition of chest width and rump width in S4 includes the following steps: extracting corresponding candidate point clouds based on point-by-point labels of the chest and rump, calculating the quantile bandwidth along the width axis in the corresponding cross section or local point cloud, and using it as chest width or rump width; chest width and rump width represent the forequarter width and hindquarter width respectively, which together reflect the lateral development status of the goat's trunk, with rump width participating in weight estimation as a hindquarter width phenotypic parameter.

6. The method for non-contact body size and weight estimation of goats based on enhanced point cloud segmentation according to claim 5, characterized in that: The central section is determined by one or more of the following: linear support vector machine, part boundary projection, or body axis geometric constraints. The linear support vector machine divides the front and back samples by the projection position of the candidate point in the body axis direction and outputs the normal vector of the candidate cutting surface.

7. The method for non-contact body size and weight estimation of goats based on enhanced point cloud segmentation according to claim 6, characterized in that, The body length is obtained by the following steps: extracting eight key points distributed along the back direction from the posterior edge of the head and neck, the anterior part of the chest, the middle part of the chest, the thoracoabdominal transition zone, the middle part of the abdomen, the posterior part of the abdomen, the anterior part of the buttocks, and the posterior edge of the buttocks; fitting a three-dimensional B-spline curve to the eight key points; and calculating the body length based on the arc length of the three-dimensional B-spline curve.

8. The method for non-contact body size and weight estimation of goats based on enhanced point cloud segmentation according to claim 7, characterized in that, Obtaining height includes the following steps: The horizontal support surface is determined based on the lowest quantile of the limb tip cloud. The key measurement point is determined by the highest point of the chest or scapular region, and the vertical distance from the key measurement point to the horizontal support surface is calculated as the body height.

9. The method for non-contact body size and weight estimation of goats based on enhanced point cloud segmentation according to claim 8, characterized in that: The training and evaluation of the GKA-enhanced point cloud segmentation network employ one or more of the following metrics: Focal Loss, mean Intersection over Union (mIoU), overall accuracy (OA), and mean class accuracy (mAcc). , , ;in, Let c be the number of points that were correctly segmented in class c. The number of points that were incorrectly classified into class c. This represents the number of points in class c that were not correctly identified. The total number of categories, This represents the total number of points in the entire point cloud.

10. A non-contact goat body size and weight estimation system based on enhanced point cloud segmentation, applied to the non-contact goat body size and weight estimation method based on enhanced point cloud segmentation as described in any one of claims 1-9, characterized in that, include: The module includes a point cloud acquisition module, a preprocessing module, a GKA-enhanced point cloud segmentation module, a body size measurement module, a volume and surface area calculation module, a weight estimation module, a quality control module, and a result output module; among which: The point cloud acquisition module is connected to the preprocessing module to acquire the three-dimensional point cloud data of the target goat; The preprocessing module is connected to the point cloud acquisition module and the GKA enhanced point cloud segmentation module. It is used to remove outliers, normalize coordinates, and sample 3D point cloud data to obtain a standardized point cloud. The GKA enhanced point cloud segmentation module is connected to the preprocessing module, body size measurement module, and volume and surface area calculation module. It is used to output point-by-point part labels for the head and neck, chest, abdomen, buttocks, left forelimb end, right forelimb end, left hindlimb end, and right hindlimb end to obtain segmented point clouds. The body size measurement module is connected to the GKA enhanced point cloud segmentation module and the weight estimation module to obtain chest circumference, abdominal circumference, chest width, hip width, body length and body height; The volume and surface area calculation module is connected to the GKA enhanced point cloud segmentation module and the weight estimation module to calculate the total volume, local volume and surface area. The weight estimation module is connected to the body size measurement module, volume and surface area calculation module, quality control module and result output module, and is used to construct morphological feature vectors and regress to output weight estimation values; The quality control module is connected to the preprocessing module, the GKA-enhanced point cloud segmentation module, the body size measurement module, the volume and surface area calculation module, the weight estimation module, and the result output module. It is used to evaluate the integrity of the point cloud, the segmentation quality, and the rationality of the body size, and output scores and low confidence prompts. The results output module is connected to the weight estimation module and the quality control module to synchronously output weight estimates, body size parameters, various scores, and visualization results.