Livestock body size measuring point extraction method, system and device and storage medium

Abstract

The invention provides a livestock body size measuring point extraction method and system, and belongs to the field of computer vision, and the method comprises the steps: obtaining single-side point cloud data of a plurality of cattle, marking the body size key region of the single-side point cloud of each cattle, and obtaining a labeled point cloud data set; the method comprises the following steps: improving existing PointNet + +, introducing deep separable convolution and residual connection operation, adding a parallelized patch attention sensing module and the like to obtain improved PointNet + +, training the improved PointNet + + through a labeled point cloud data set to obtain a cattle body size key area identification and positioning model CatttlePartNet, and identifying and positioning the cattle body size key area according to the training result of the improved PointNet + + and the training result of the improved PointNet + + and the training result of the improved PointNet + + and the training result of the improved PointNet + +. And inputting the unilateral point cloud data of the to-be-measured cattle into the cattle body size key area identification and positioning model CattlePartNet, identifying different body size key areas of the cattle, and performing calculation according to the point set of the body size key areas to obtain body size parameters. According to the method, the key area and the measuring point of the cattle body can be accurately identified, so that accurate body size parameters are obtained, and the measuring precision and stability are improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of livestock body measurement, and particularly relates to a method, system, device and storage medium for extracting livestock body measurement points. Background Art

[0002] Body measurement parameters of beef cattle, such as body height, body slant length, chest girth and abdominal girth, etc., reflect the body size and growth and development status of cattle, and are closely related to the genetic breeding, population reproductive performance and production performance of the cattle herd. They are important reference indicators for measuring the breeding effect and the degree of large-scale and standardized feeding. At the same time, the breeder can estimate the weight of the cattle by measuring these body measurement parameters. In order to evaluate the healthy reproduction and genetic breeding effect of cattle, the breeder must regularly measure the body measurement and weight of cattle to master their health and nutritional status. However, the traditional method for measuring the body size of beef cattle usually requires driving the cattle to a restraining rack and using tools such as a tape measure and a measuring rod for manual contact measurement. This method not only easily causes stress response in cattle, but also is time-consuming, laborious and inefficient in the measurement process, resulting in greater measurement difficulty and error. Currently, the mainstream method for automatically calculating cattle body measurement data is based on two-dimensional images and is achieved by restricting the posture of cattle through a fixing device.

[0003] For example, the Chinese invention patent with the authorization announcement number CN115633955B discloses a dairy cattle body measurement system, including a restraining rack and a measuring machine installed on the restraining rack, which can quickly measure the body size of dairy cattle, and the measurement steps are convenient and fast. Another example is the Chinese invention patent with the authorization announcement number CN105726028B, a non-contact measurement method for Xinjiang brown cattle body measurement indexes, which sets up a set of measurement system including a measurement channel, five groups of cameras, a host computer, a binocular camera calibration system and an image acquisition system, and the image measurement system processes the image information and then displays the measurement results.

[0004] Although the above methods can achieve the body measurement of cattle, this method requires driving the cattle to a restraining rack and then setting up multiple sets of cameras to collect the image data of the cattle. This method has strict requirements for the posture of cattle and the acquisition environment, and at the same time, it is impossible to obtain the indexes related to the body surface shape of cattle. Summary of the Invention

[0005] In order to solve the above problems, the present invention provides a method for extracting livestock body measurement points.

[0006] In order to achieve the above object, the present invention provides the following technical solutions:

[0007] A method for extracting livestock body measurement points, comprising:

[0008] Obtaining the unilateral point cloud data of multiple cattle, and marking the body measurement key areas of the unilateral point cloud of each cattle to obtain a labeled point cloud data set;

[0009] Using the PointNet++ deep learning model for point clouds as the backbone network, a residual connection MLP module is added at the very front of the backbone network. After the set abstraction SA module of the backbone network, a DSC-MLP module composed of depthwise separable convolution and residual connection and a perceptual patch attention module are added. The perceptual patch attention module includes a parallelized patch perception attention PPA module, a dimension perception selective integration DASI module, and a multi-expansion channel refinement MDCR module. The Adam optimizer of the backbone network is replaced with the Sophia optimizer to obtain the improved PointNet++; the improved PointNet++ is trained with a labeled point cloud dataset of cattle to obtain the cattle body measurement key area recognition and localization model CattlePartNet;

[0010] Input the unilateral point cloud data of the cattle to be measured into the cattle body measurement key area recognition and localization model CattlePartNet to identify different body measurement key areas of the cattle;

[0011] Calculate the mean curvature H and Gaussian curvature K of the point sets of different body measurement key areas of the cattle respectively, and determine the candidate areas for cattle body measurement points by judging the curvature value combination; according to the extracted candidate areas, calculate the mean curvature of each area, and obtain the body measurement parameters based on the mean curvature.

[0012] Preferably, the key body measurement areas of each unilateral point cloud of the cattle are labeled to obtain a labeled point cloud dataset, specifically:

[0013] Filter, denoise, cluster and segment, and surface reconstruct the unilateral point cloud data of the cattle, and remove the wall and window noise points in the scene point cloud data to obtain the preprocessed unilateral point cloud data of beef cattle;

[0014] Enhance and expand the preprocessed unilateral point cloud data of beef cattle, and then use a point cloud annotation tool to label different key body measurement areas of each cattle point cloud to obtain a labeled point cloud dataset.

[0015] Preferably, the different key body measurement areas include the withers area, the scapular area, and the ischium area, and the body measurement parameters include the positions of the body height measurement point, the shoulder end point, and the ischium node.

[0016] Preferably, the PA module contains 3 parallel branches: a local branch, a global branch, and a serial convolution branch. The processing of data by the PA module is specifically:

[0017] Given an input feature tensor First, obtain the feature tensor after adjusting the number of channels through pointwise convolution Then, through the local branch, the global branch, and the serial convolution branch, calculate the local branch convolution result respectively Global branch convolution result and serial branch convolution result Add the three results to obtain the convolution feature fusion result of the multi-branch

[0018] Preferably, the spatial Euclidean distance formula is used to calculate the linear body dimensions of body height, body diagonal length, and hip height.

[0019] Preferably, a lidar camera is used to collect the unilateral point cloud data of cattle.

[0020] The present invention also provides a livestock body dimension measurement point extraction method system, including:

[0021] A training data production module, configured to obtain multiple unilateral point cloud data of cattle, and label the body dimension key areas of each unilateral point cloud of cattle to obtain a labeled point cloud data set;

[0022] A model construction module, configured to use the point cloud deep learning model PointNet++ as the backbone network, add a residual connection MLP module in front of the backbone network, add a DSC-MLP module composed of depthwise separable convolution and residual connection and a perceptual patch attention module after the set abstraction SA module of the backbone network. The perceptual patch attention module includes a parallelized patch perceptual attention PPA module, a dimension-aware selective integration DASI module, and a multi-expansion channel refinement MDCR module. Replace the Adam optimizer of the backbone network with the Sophia optimizer to obtain an improved PointNet++; train the improved PointNet++ through the labeled point cloud data set to obtain a cattle body dimension key area recognition and localization model CattlePartNett;

[0023] A key area extraction module, configured to input the unilateral point cloud data of the cattle to be measured into the cattle body dimension key area recognition and localization model CattlePartNet to identify different body dimension key areas of the cattle;

[0024] A measurement point extraction module, configured to calculate the mean curvature H and Gaussian curvature K of the point sets of different body dimension key areas of the cattle respectively, and determine the candidate areas of the cattle body dimension measurement points by judging the curvature value combination; according to the extracted candidate areas, calculate the mean curvature of each area, and obtain the body dimension parameters according to the mean curvature.

[0025] The present invention also provides a computer device, including a memory, a processor, and a computer program stored on the memory. The processor executes the computer program to implement the steps of any one of the livestock body dimension measurement point extraction methods.

[0026] The present invention also provides a computer-readable storage medium, on which a computer program is stored. When the computer program is loaded by a processor, it can execute the steps described in any one of the methods for extracting body measurement points of livestock.

[0027] The method for extracting body measurement points of livestock provided by the present invention has the following beneficial effects:

[0028] By collecting unilateral point cloud data of cattle instead of camera images, the process of collecting point cloud data does not require specific environmental constraints and can be directly completed in the natural scene where cattle move. There is no need to drive the cattle to a restraining rack or forcibly fix the cattle. The requirements for the posture of cattle and the collection environment are low, which reduces the requirements for the measurement site and equipment, improves the flexibility and convenience of practical applications, and reduces the stress response and behavioral interference of cattle. Through the improvement of the PointNet++ model, including residual connection, depthwise separable convolution, and perceptual patch attention module, it is possible to first identify different key body measurement regions of cattle, and then calculate the mean curvature H and Gaussian curvature K of the point sets in different key body measurement regions of cattle respectively. The candidate regions for body measurement points of cattle are determined by judging the combination of curvature values. According to the extracted candidate regions, the mean curvature of each region is calculated, and the body measurement parameters are obtained based on the mean curvature, so that the present invention can accurately identify the key regions and measurement points of the cattle body, and then obtain accurate body measurement parameters, improving the measurement accuracy and stability. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] In order to more clearly illustrate the embodiments of the present invention and their design schemes, the accompanying drawings required for the present embodiments will be briefly introduced below. The accompanying drawings in the following description are only partial embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0030] Figure 1 is a flowchart of the method for extracting body measurement points of livestock in Embodiment 1 of the present invention;

[0031] Figure 2 is the structural diagram of the CattlePartNet model proposed by the present invention;

[0032] Figure 3 is the segmentation result diagram of the key body measurement regions of beef cattle; where, Figure 3 in (a) is a set of feature parts manually labeled; Figure 3 in (b) is a set of feature parts labeled by the model;

[0033] Figure 4 is the visualization of the body measurement results of three cattle. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0034] To enable those skilled in the art to better understand the technical solution of the present invention and be able to implement it, the present invention will be described in detail below in conjunction with the accompanying drawings and specific embodiments. The following embodiments are only used to more clearly illustrate the technical solution of the present invention and cannot be used to limit the protection scope of the present invention.

[0035] Embodiment 1

[0036] The present invention provides a method for extracting body measurement points of livestock, specifically as Figure 1 shown, including the following steps:

[0037] Step 1: Use a lidar camera to collect the unilateral point cloud data of cattle, and perform preprocessing operations such as filtering, denoising, clustering segmentation, and surface reconstruction on the point cloud data to remove noise points such as walls and windows in the scene point cloud data, and obtain a complete and clear unilateral point cloud of beef cattle to ensure the quality and accuracy of the data.

[0038] Step 2: According to the point cloud data obtained in Step 1, enhance and expand the data, and then use a point cloud annotation tool to annotate 3 key body measurement regions (withers region, scapular region, ischial region) of each cattle point cloud to obtain a point cloud data set with labels already made.

[0039] Step 3: Aiming at the point cloud neighborhood relationship and spatial geometric distribution problems of PointNet++, depthwise separable convolution and residual connection operations are introduced into the network, and a parallelized perceptual patch attention module is added to fully consider the importance of neighborhood information and spatial distribution information. Aiming at the problem that the network depth is relatively shallow, resulting in a small receptive field of points, a method of deepening the network layer and scaling the receptive field is adopted, and the Sophia optimizer is introduced to construct a new model CattlePartNet dedicated to the recognition and positioning of key body measurement regions of cattle, Figure 2 shows the improved network architecture diagram, which is used to complete the segmentation task of the characteristic parts of the beef cattle point cloud. The red frame line indicates the difference from the original model. Then, the point cloud data set obtained in Step 2 is used to train the CattlePartNet model, and a deep learning model for the recognition of key body measurement regions of cattle is trained.

[0040] Specifically, the process of improving the existing point cloud deep learning model PointNet++ in Step 3 includes:

[0041] Step 31: Using the point cloud deep learning model PointNet++ as the backbone network, since the ball query algorithm is used to find the points in the neighborhood in the original model, considering that the query radius of the ball query is specific to the data set, its query radius is fixed and can have a significant impact on performance. As shown in Equation x m a+1 =R n:(m,n)∈N {h Θ ([xn a ; p n a -p m a )} in the relative coordinates (Δp = p n a -p m a ) makes network optimization more difficult and leads to performance degradation. Without normalization, the values of relative positions (Δp = p n a -p m a ) are quite small (less than the radius), and the network needs to learn larger weights to apply to Δp. This makes optimization more difficult, especially since weight decay is mainly used to reduce the weights of the network, thus ignoring the influence of relative positions. Therefore, normalize the relative positions (Δp normalization), that is, x m a+1 = R n:(m,n)∈N {h Θ ([x n a ; p n a -p m a )}, dividing the relative positions by the neighborhood query radius. Adopt the normalization method for scaling to alleviate weight decay, which is mainly used to reduce the weights of the network to make its range more appropriate, and at the same time reduce the variance of Δp between different stages.

[0042] Step 32: A new module named DSC-MLP (Depthwise Separable Convolution Multilayer Perceptron) is introduced into the point cloud deep learning model PointNet++. This module combines depthwise separable convolution and residual connections, aiming to efficiently extract point cloud features. This module inherits the classic grouping structure in the SetAbstraction (SA) module of PointNet++. The difference is that, aiming at the point cloud neighborhood relationship and spatial geometric distribution problems of PointNet++, a direct connection, that is, a residual connection operation, is added between the input end and the output end of the model. DSC-MLP consists of depthwise separable convolution and residual connection operations, fully considering the importance of neighborhood information and spatial distribution information to achieve efficient and effective model expansion. The network is named CattlePartNet and is used to complete the task of segmenting the feature parts of beef cattle point clouds.

[0043] Next, compared with the backbone network PointNet++, CattlePartNet adds a layer of mixed feature vector MLP at the forefront of the network structure to increase the dimension of the data. And the SA module is applied to start local feature extraction. Then it enters the DSC-MLP module, which deepens the depth of the network by using depthwise separable convolution and residual connection operations. The dimensions of the 4 MLPs first increase and then decrease, forming an inverted bottleneck structure, and the dimension of the point cloud data does not change before and after passing through the inverted residual block. After that is the Feature Propagation (FP) module inherited from the original model, which performs inverse distance weighted interpolation on the incoming data and performs skip connection splicing with the data lost in the Reduction.

[0044] Finally, it enters the fully connected layer to complete point cloud segmentation and output the classification result of each point. At the same time, the PPA (Parallelized Patch-Aware Attention) module is introduced. The PPA replaces the traditional convolution operations in the basic components of the encoder and decoder to more effectively solve the problem of losing necessary information during the repeated downsampling process. The main advantage of the PPA is its multi-branch feature extraction strategy. The PPA adopts a parallel multi-branch method, and each branch extracts features of different scales and levels. This strategy effectively captures multiple multi-scale features of the target and improves the accuracy of small object detection.

[0045] Specifically, the PPA module contains 3 parallel branches: local, global, and serial convolution branches. Given an input feature tensor First, a feature tensor with adjusted number of channels is obtained through pointwise convolution Then, through three branches, the convolution results of the local branch the convolution result of the global branch and the convolution result of the serial branch are calculated respectively. The three results are added to obtain the convolution feature fusion result of the multi-branch After multi-branch feature extraction, adaptive feature enhancement is achieved through the attention mechanism. The attention module consists of a series of efficient channel attention and spatial attention components. is obtained through the one-dimensional channel attention map and the two-dimensional channel attention map The process is as follows: F” = δ(β(dropout(F s ))), where represents element-wise multiplication, and They respectively represent the results after channel and spatial feature selection. δ(·) and β(·) respectively represent Relu and regularization. Finally, the result of the PPA module is output, that is, the output feature tensor after multi-branch feature extraction, feature fusion, attention mechanism enhancement, regularization, and activation function processing.

[0046] The present invention uses the point cloud deep learning model PointNet++ as the backbone network, and introduces depthwise separable MLPs operations into the network, that is, dividing the fully connected layer into channel-wise convolution and point-wise convolution, which can reduce the computational amount and thus strengthen the extraction of point-wise features. Although all three layers of MLP in the original SA module are calculated on neighborhood features, DFS-MLP separates the MLP into a single layer calculated on neighborhood features (between the grouping layer and the Reduction layer) and two layers of point features (after the Reduction layer). A residual connection operation is introduced, that is, a residual connection is added between the input and output of the PointNet++ model to achieve the purpose of alleviating the vanishing gradient. In this way, the network can not only transmit information through the forward propagation path, but also directly transmit the original input information to subsequent layers through the residual connection. A parallel patch-aware attention module is added. Specifically, this module includes a parallel patch-aware attention (PPA) module, a dimension-aware selective integration (DASI) module, and a multi-expanded channel refinement (MDCR) module. The PPA module adopts a multi-branch feature extraction strategy to capture feature information at different scales and levels. The DAS module supports adaptive channel Q selection and fusion. The MDCR module captures spatial features in different receptive field ranges through multiple depthwise separable convolutional layers, and enhances the feature extraction ability of the model through this parallel patch-aware attention mechanism. And the Sophia optimizer is introduced. The original model uses the Adam optimizer, but its complex second-order (Hessian-based) optimizer usually leads to excessive overhead per step. The Sophia optimizer uses a lightweight estimate of the diagonal Hessian as a precondition. The update is the moving average of the gradient divided by the moving average of the estimated Hessian, followed by element-wise clipping. This clipping controls the worst-case update size and suppresses the negative effects of non-convexity and rapid changes in the Hessian along the trajectory. Finally, the improved PointNet++ is obtained.

[0047] Step 4: Input the unilateral point cloud data of the cattle with labels (the withers region, scapular region, and ischial region have been labeled) into the new model CattlePartNet to identify three key body size regions of the cattle.

[0048] The withers region is the area where the body height measurement point is located, the scapular region is the area where the shoulder endpoint is located, the ischial region is the area where the ischial node is located, and the Euclidean distance between the shoulder endpoint and the ischial node is the body diagonal length of the beef cattle; since both the body height and the hip height are perpendicular to the ground, they can be obtained through the vertical distances from the body height measurement point and the ischial node to the horizontal ground.

[0049] To improve the positioning efficiency of the body measurement points of beef cattle, the candidate regions of the feature points of three key body measurement parts are extracted first. The surface description method using the mean curvature H and the Gaussian curvature K is adopted to extract the candidate regions of the body measurement points, so as to reduce the subsequent data processing steps.

[0050] By calculating the mean curvature and the Gaussian curvature, according to the concave and convex conditions of the region where the body measurement points are located on the point cloud surface, when the curvature value combination of the point is H = 0, K > 0, this combination is determined as the shape of the region where the body measurement points of the cattle are located, and the cattle body point cloud is classified to obtain the candidate regions of the body measurement points.

[0051] For the contour point sets of each body measurement feature part obtained, the mean curvature of each point is compared. In the feature region, the point with the maximum mean curvature is determined as the body measurement point sought by the present invention. Among them, the withers region determines the location of the body height measurement point, the scapular region determines the location of the shoulder endpoint, and the ischial region determines the location of the ischial node.

[0052] For the three body measurement points obtained (body height measurement point, shoulder endpoint, ischial node), the spatial Euclidean distance formula is used to calculate three linear body measurements including body height, body diagonal length, and hip height. Figure 3 It is the visualization of the body measurement results of three cattle. Among them, Figure 3 (a) of is a set of feature parts manually marked; Figure 3 (b) of is a set of feature parts marked by the model.

[0053] Step 5: According to the key body measurement regions of the cattle obtained in Step 4, by calculating the mean curvature and the Gaussian curvature of each point in the key body measurement regions, judge the concave and convex conditions of the region where the body measurement points are located. When the curvature value combination of the point is H = 0, K > 0, this combination is determined as the shape of the region where the body measurement points of the cattle are located, and the cattle body point cloud is classified to obtain the candidate regions of the body measurement points. For the candidate region point sets of each body measurement feature part obtained, the mean curvature of each point is compared. In the candidate point set of the key points, the point with the maximum mean curvature is determined as the body measurement point sought by the present invention. Among them, the withers region determines the location of the body height measurement point, the scapular region determines the location of the shoulder endpoint, and the ischial region determines the location of the ischial node. For the three body measurement points obtained (body height measurement point, shoulder endpoint, ischial node), the spatial Euclidean distance formula is used to calculate three linear body measurements including body height, body diagonal length, and hip height.

[0054] The method for extracting livestock body measurement points provided by the present invention has the following beneficial effects:

[0055] 1. Non-contact measurement, avoiding stress reactions to cattle. The present invention utilizes point cloud data and deep learning technology to achieve automatic extraction of cattle body measurement points, without the need for cattle farm personnel to drive the cattle to a restraining rack, nor the need to forcibly fix the cattle, reducing the stress reactions and behavioral interference of the cattle.

[0056] 2. Improving measurement efficiency. Traditional methods require manual operation or the assistance of mechanical devices to fix the cattle, which is time-consuming and requires high requirements for operators. The present invention realizes efficient and rapid extraction of body measurement points and data calculation through an improved point cloud component segmentation model and an automated algorithm, shortening the measurement time.

[0057] 3. Improving measurement accuracy and stability. By introducing an improved PointNet++ model, including residual connections, depthwise separable convolutions, and a perceptual patch attention module, the present invention can accurately identify the key areas and measurement points of the cattle body, reducing the errors that may be brought by manual measurement.

[0058] 4. Ability to adapt to complex scenarios. The point cloud data acquisition process does not require specific environmental constraints and can be directly completed in the natural scenarios where the cattle herd is active, reducing the requirements for the measurement site and equipment, and improving the flexibility and convenience of practical applications.

[0059] In summary, the technical solution of the present invention avoids the necessity of fixing the cattle posture in traditional methods through non-contact automated measurement technology, and at the same time has relatively significant advantages in terms of efficiency, accuracy, and adaptability of the application scenario.

[0060] The present invention also provides a system for the method of extracting livestock body measurement points, including:

[0061] A training data production module, used to obtain the unilateral point cloud data of multiple cattle, and label the key body measurement areas of each unilateral point cloud of the cattle to obtain a labeled point cloud data set.

[0062] The model construction module is used to take the PointNet++ deep learning model for point clouds as the backbone network, add a residual connection MLP module at the very front of the backbone network, add a DSC-MLP module composed of depthwise separable convolution and residual connection and a perceptual patch attention module after the set abstraction SA module of the backbone network. The perceptual patch attention module includes a parallelized patch perceptual attention PPA module, a dimension-aware selective integration DASI module, and a multi-expansion channel refinement MDCR module. Replace the Adam optimizer of the backbone network with the Sophia optimizer to obtain the improved PointNet++; train the improved PointNet++ with the labeled point cloud dataset to obtain the cattle body measurement key area recognition and localization model CattlePartNett.

[0063] The key area extraction module is used to input the unilateral point cloud data of the cattle to be measured into the cattle body measurement key area recognition and localization model CattlePartNet to identify different body measurement key areas of the cattle.

[0064] The measurement point extraction module is used to calculate the mean curvature H and Gaussian curvature K of the point sets of different body measurement key areas of the cattle respectively, determine the candidate areas of the cattle body measurement points by judging the curvature value combination; calculate the mean curvature of each area according to the extracted candidate areas, and obtain the body measurement parameters according to the mean curvature.

[0065] Each module in the above livestock body measurement point extraction method system can be implemented in whole or in part by software, hardware and their combination. The above modules can be embedded in the processor of the computer device in hardware form or independent of it, or stored in the memory of the computer device in software form, so that the processor can call and execute the operations corresponding to the above modules.

[0066] The present invention also provides a computer device, including a memory, a processor and a computer program stored on the memory. The processor executes the computer program to implement the steps in the embodiment of the livestock body measurement point extraction method. The specific implementation method can refer to the method embodiment and will not be elaborated here.

[0067] Furthermore, the present invention also provides a non-transitory computer-readable storage medium containing instructions, and a computer program is stored on the storage medium. For example, a memory containing instructions, and the above instructions can be executed by the processor of the computer device to complete the above method. For example, the non-transitory computer-readable storage medium can be ROM, random access memory (RAM), CD-ROM, magnetic tape, floppy disk and optical data storage device, etc. When the computer program is executed by the processor, it can implement the steps in the embodiment of the livestock body measurement point extraction method. The specific implementation method can refer to the method embodiment and will not be elaborated here.

[0068] Those skilled in the art should understand that the embodiments of the present invention may provide a method, a system or a computer program product. Therefore, the present invention may take the form of an entirely hardware embodiment, an entirely software embodiment or an embodiment combining software and hardware aspects. Moreover, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) that contain computer-usable program code.

[0069] The present invention is described with reference to the flowcharts and / or block diagrams of methods, apparatuses (systems) and computer program products according to the embodiments of the present invention. It should be understood that each flow and / or block in the flowchart and / or block diagram can be implemented by computer program instructions, and the combination of flows and / or blocks in the flowchart and / or block diagram can also be implemented by computer program instructions. These computer program instructions can be provided to the processor of a general-purpose computer, a special-purpose computer, an embedded processor or other programmable data processing devices to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing devices generate a device for implementing the functions specified in Figure 1 one flow or multiple flows and / or blocks Figure 1 one block or multiple blocks.

[0070] These computer program instructions can also be stored in a computer-readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer-readable memory generate a manufactured article including an instruction device, and the instruction device implements the functions specified in Figure 1 one flow or multiple flows and / or blocks Figure 1 one block or multiple blocks.

[0071] These computer program instructions can also be loaded onto a computer or other programmable data processing device, so that a series of operation steps are executed on the computer or other programmable device to generate a computer-implemented process, and thus the instructions executed on the computer or other programmable device provide steps for implementing the functions specified in Figure 1 one flow or multiple flows and / or blocks Figure 1 one block or multiple blocks.

[0072] It should be noted that the above-described specific embodiments can enable those skilled in the art to more comprehensively understand the present invention, but do not limit the present invention in any way. Therefore, although the present specification and embodiments have described the present invention in detail, those skilled in the art should understand that the present invention can still be modified or equivalently replaced; and all technical solutions and improvements that do not depart from the spirit and scope of the present invention are covered by the protection scope of the patent of the present invention. Any reference signs in the claims should not be construed as limiting the claimed claim. Any simple changes or equivalent replacements of technical solutions that can be obviously obtained by any person skilled in the art within the technical scope disclosed by the present invention all belong to the protection scope of the present invention.

Claims

1. A method for extracting livestock body measurement points, characterized in that: include: Obtain unilateral point cloud data of multiple cattle, and mark the key body size areas of the unilateral point cloud of each cattle to obtain a labeled point cloud dataset; The point cloud deep learning model PointNet++ is used as the backbone network, a residual connection MLP module is added to the front of the backbone network, and a DSC-MLP module composed of deep separable convolution and residual connection and a perceptual patch attention module are added after the set abstraction SA module of the backbone network. The perceptual patch attention module includes a parallelized patch-aware attention PPA module, a dimension-aware selective integration DASI module and a multi-extended channel refinement MDCR module. The Adam optimizer of the backbone network is replaced with the Sophia optimizer to obtain an improved PointNet++. The improved PointNet++ is trained with a labeled point cloud dataset to obtain the CattlePartNet model for identifying and locating key areas of cattle body size. The unilateral point cloud data of the cattle to be tested is input into the cattle body key area recognition and positioning model CattlePartNet to identify the different body key areas of the cattle; The average curvature H and Gaussian curvature K of the key area points of different body sizes of cattle are calculated respectively, and the candidate areas of cattle body size measurement points are determined by judging the combination of curvature values; based on the extracted candidate areas, the average curvature of each area is calculated, and the body size parameters are obtained according to the average curvature.

2. The method for extracting livestock body size measurement points according to claim 1, characterized in that: The key area of ​​the body size of each cow's unilateral point cloud is annotated to obtain a labeled point cloud dataset, specifically: The unilateral point cloud data of cattle is filtered, denoised, clustered, and reconstructed to remove the wall and window noise in the scene point cloud data, and the preprocessed unilateral point cloud data of beef cattle is obtained; The preprocessed unilateral point cloud data of beef cattle are enhanced and expanded, and then the point cloud annotation tool is used to annotate the key areas of different body sizes of each cattle point cloud to obtain a labeled point cloud dataset.

3. The method for extracting livestock body size measurement points according to claim 2, characterized in that: The different key areas of body size include the withers area, the scapular area and the ischium area, and the body size parameters include the positions of the body height measurement point, the shoulder endpoint and the ischium node.

4. The method for extracting livestock body size measurement points according to claim 1, characterized in that: The PA module includes three parallel branches: a local branch, a global branch, and a serial convolution branch. The PA module processes data specifically as follows: Given an input feature tensor First, the feature tensor after adjusting the number of channels is obtained through point-by-point convolution Then, the local branch convolution results are calculated through the local branch, global branch and serial convolution branch respectively. Global branch convolution results And the serial branch convolution results Add the three results to get the multi-branch convolution feature fusion result 5. The method for extracting livestock body size measurement points according to claim 3, characterized in that: The linear body dimensions of body height, body oblique length and hip height were calculated using the spatial Euclidean distance formula.

6. The method for extracting livestock body size measurement points according to claim 1, characterized in that: A lidar camera is used to collect point cloud data of one side of the cow.

7. A livestock body measurement point extraction method system, characterized in that: include: The training data production module is used to obtain the unilateral point cloud data of multiple cattle and mark the key body size areas of the unilateral point cloud of each cattle to obtain a labeled point cloud data set; A model building module is used to use the point cloud deep learning model PointNet++ as the backbone network, add a residual connection MLP module at the front of the backbone network, add a DSC-MLP module composed of deep separable convolution and residual connection and a perception patch attention module after the set abstraction SA module of the backbone network, the perception patch attention module includes a parallelized patch-aware attention PPA module, a dimension-aware selective integration DASI module and a multi-extended channel refinement MDCR module, and replace the Adam optimizer of the backbone network with the Sophia optimizer to obtain an improved PointNet++; the improved PointNet++ is trained by using a labeled point cloud dataset to obtain a cattle body size key area recognition and positioning model CattlePartNett; The key area extraction module is used to input the single-side point cloud data of the cattle to be tested into the cattle body size key area identification and positioning model CattlePartNet to identify the different body size key areas of the cattle; The measurement point extraction module is used to calculate the average curvature H and Gaussian curvature K of the point sets in different key areas of the cattle's body size, and determine the candidate areas of the cattle's body size measurement points by judging the combination of curvature values; based on the extracted candidate areas, the average curvature of each area is calculated, and the body size parameters are obtained based on the average curvature.

8. A computer device comprising a memory, a processor and a computer program stored in the memory, characterized in that: The processor executes the computer program to implement the steps of the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is loaded into a processor, it can execute the steps of the method according to any one of claims 1 to 6.

Images