Pig important growth character prediction method based on multi-source fusion

Abstract

The invention discloses a multi-source fusion-based pig important growth trait prediction method, and belongs to the technical field of animal growth trait prediction. The objective of the invention is to solve the problems of long time consumption and poor measurement result accuracy of the existing method for obtaining important growth traits such as backfat thickness and eye muscle area. According to the method, body size parameters and morphological characteristics of pigs are obtained based on point cloud data corresponding to the pigs, then body size parameter characteristics are adopted to enhance the morphological characteristics to obtain enhanced morphological characteristics, the enhanced morphological characteristics are sent to a Flaten layer to be flattened, and a second characteristic vector is obtained; obtaining a first feature vector based on the pig breed type feature vector, the growth stage feature vector, the slaughter day age and the exercise amount corresponding to the live pig; and the first feature vector and the second feature vector are spliced and then are sent to a neural network model to obtain a predicted output important growth trait predicted value.

Description

Technical Field

[0001] This invention belongs to the field of animal growth trait prediction technology, specifically relating to a method for predicting important growth traits in pigs. Background Technology

[0002] In pig production, growth traits are important breeding traits and key indicators for measuring pig meat production capacity. Backfat thickness and eye muscle area are two particularly important growth traits, not only key indicators of pork quality, feeding efficiency, and lean meat percentage, but also crucial factors affecting sow reproductive performance, holding significant importance for the livestock industry, meat processing industry, and consumers. Currently, these traits are mainly obtained through contact-based ultrasonic measurement methods. However, the accuracy of this method is greatly affected by the operator's skill and the pig's condition, posing challenges for large-scale application and potentially causing stress reactions and biosecurity risks in pigs. With the pig farming industry developing towards large-scale and intelligent operations, there is an urgent need for a simple and efficient non-contact measurement method.

[0003] Early research on body size measurement primarily utilized CCD (Charge-coupled Device) cameras to acquire images of livestock from above, side, or rear. These acquired data were all two-dimensional images. Image processing techniques were then used to detect feature points in these two-dimensional images before measuring body length, height, and width. In the 1990s, international researchers began using computer vision for non-contact body size measurement in pigs and cattle. Minagawa first proposed using image processing techniques for pig body size measurement. He used a CCD camera to acquire RGB images of the pig's back, applied a fixed threshold technique to segment the image to obtain a clear outline of the pig, and finally obtained the actual length of each pixel using a calibration object. The body size length was then calculated based on the number of pixels at the corresponding position. Subsequently, numerous studies on 2D image-based body size measurement emerged in cattle and pigs. Stajnko et al. used a thermal imager to acquire two-dimensional images of the bull's abdomen, using these images to measure body width and hip height. Thermal images made it easier to extract the target area of ​​the bull compared to conventional RGB images. Tasdemir et al. used four cameras to simultaneously acquire RGB images of cows from different perspectives to measure body size parameters such as shoulder height, hip height, body length, and hip width. The accuracy of all body size parameters was over 95%.

[0004] Two-dimensional image-based computer vision measurement methods are easily affected by lighting and surrounding background during image acquisition, and can only measure two-dimensional body size parameters, such as height, length, and width, but cannot measure three-dimensional body size parameters such as chest circumference and abdominal circumference, which are crucial for livestock body condition analysis. With the introduction of consumer-grade depth cameras by companies such as Microsoft, depth image-based livestock body size measurement methods have become a research hotspot. Salau developed an extrinsic parameter calibration algorithm for six Kinect cameras based on a wooden cube, and realized three-dimensional reconstruction of dairy cows based on the extrinsic parameter matrix. Based on the results of this three-dimensional reconstruction, the authors completed the measurement of cow teat length and ischial tuberosity height using a manual labeling method. They also achieved automated measurement of hind leg angle and teat height from the ground based on the morphological characteristics of the cow's udder and hind legs. Pezzuolo et al. used four KinectV1 cameras to acquire depth images of the sides of the abdomen, head, and back of dairy cows, and completed three-dimensional reconstruction of the cows based on hemispherical geometric calibration objects. By using threshold settings and filters, they automatically identified specific body parts and calculated various body size parameters accordingly.

[0005] While these methods have yielded some results, they all focus on obtaining three-dimensional body size parameters. Existing methods cannot directly obtain data on important growth traits such as backfat thickness and eye muscle area. Backfat thickness is a key indicator of fat deposition in pigs, and eye muscle area is a key indicator of muscle (lean meat) development; therefore, these two indicators are particularly crucial. Since backfat thickness refers to the thickness of subcutaneous fat on the pig's back, it is usually measured at specific locations using ultrasound equipment after slaughter or while the pig is still alive. However, this method is not only cumbersome but also time-consuming, especially with ultrasound equipment, which consumes a significant amount of measurement time. Furthermore, the constant movement of the pig leads to substantial measurement errors. Current methods for obtaining three-dimensional body size parameters cannot accurately obtain these indicators. Similarly, eye muscle area, which refers to the cross-sectional area of ​​the longissimus dorsi muscle, suffers from the same problem. Summary of the Invention

[0006] This invention addresses the problems of time-consuming and inaccurate measurement results in existing methods for obtaining important growth traits such as backfat thickness and eye muscle area.

[0007] Multi-source fusion-based methods for predicting important growth traits in pigs include:

[0008] The body size parameters and morphological features of pigs are obtained based on the point cloud data corresponding to the pigs. Then, the morphological features are enhanced by using the body size parameter features to obtain enhanced morphological features, which are then fed into the Flatten layer to flatten and obtain the second feature vector. The first feature vector is obtained based on the pig breed type feature vector, growth stage feature vector, slaughter age and exercise level.

[0009] The first and second feature vectors are concatenated and fed into the neural network model to obtain the predicted values ​​of important growth traits.

[0010] Furthermore, the predicted values ​​for the important growth traits include predicted values ​​for backfat thickness and eye muscle area.

[0011] Furthermore, before obtaining the body size parameters and morphological characteristics of pigs based on the point cloud data corresponding to the pigs, it is necessary to first obtain point cloud data from at least two perspectives and perform point cloud registration. After point cloud registration, the point cloud data is used to obtain the body size parameters and morphological characteristics of the pigs.

[0012] Furthermore, the point cloud registration process includes:

[0013] First, coarse registration is performed using the RANSAC algorithm or the fast global registration algorithm based on fast point feature histogram features. Then, fine registration is performed using the iterative nearest point algorithm ICP.

[0014] Furthermore, the process of obtaining the morphological features of pigs based on point cloud data adopts a PointNet or PointNet++ model with the sigmoid activation function of the last layer removed.

[0015] Furthermore, enhanced morphological features are obtained by using body size parameter features to enhance morphological features. , for transpose, This is the vector obtained by normalizing the volume scale parameters, i.e., the normalized volume scale parameter features; These are morphological features obtained based on point cloud data.

[0016] Furthermore, the exercise data used to obtain the first feature vector is the average daily effective exercise volume, which is obtained through the following steps:

[0017] Acquire acceleration data of pigs over a period of time, and calculate the median acceleration value, denoted as threshold A. Then, for each day's acceleration data, using the acceleration data corresponding to time period l as a unit, count whether there are two or more acceleration data points greater than threshold A within a unit. If so, the current unit is counted as valid motion data; otherwise, it is not considered valid motion data. Count the number N of valid motion data points in a day, and normalize them to obtain the normalized daily valid motion volume. Here, n represents n units per hour;

[0018] Finally, the average daily effective movement volume within the slaughter age period was calculated. As a measure of exercise for pigs.

[0019] Furthermore, in the process of acquiring the acceleration data of pigs over a period of time, the period is 5 days; the duration is 10 minutes.

[0020] Furthermore, the pig breed type feature vector used to obtain the first feature vector is obtained through the following steps:

[0021] Based on the pig breed, one-hot encoding is used for encoding, and then each bit of the encoding is used as an element to form a feature vector.

[0022] Furthermore, the growth stage feature vector used to obtain the first feature vector is obtained through the following steps:

[0023] Based on the growth stages of pigs, one-hot encoding is used for encoding, and then each encoded element is used as an element to form a feature vector.

[0024] Beneficial effects:

[0025] This invention collects depth image data of the back and abdomen of pigs and proposes for the first time to obtain prediction input directly from three-dimensional point cloud data. It also explores the optimal data preprocessing method and then uses the processed data to predict the backfat thickness and eye muscle area of ​​pigs. This not only provides a new approach to predicting backfat thickness and eye muscle area, but also enables rapid prediction based on daily monitoring data and achieves good prediction results. Attached Figure Description

[0026] Figure 1 This is a schematic diagram of the prediction process for important growth traits in pigs.

[0027] Figure 2 This is a list of acceleration data for a specific time period during the day.

[0028] Figure 3 This is an acceleration curve that corresponds to a period of 5 days.

[0029] Figure 4 This is a graph showing the distribution of prediction accuracy for different models in breeding value estimation. Detailed Implementation

[0030] This embodiment is a method for predicting important growth traits of pigs based on multi-source fusion. It uses computer vision technology to construct a non-contact measurement model for backfat thickness and eye muscle area of ​​pigs, providing technical support for non-contact intelligent measurement of backfat thickness and eye muscle area of ​​pigs.

[0031] Because the datasets, preprocessing techniques, metrics, and models used in the studies differ, the results obtained by comparing existing methods are not entirely reasonable. Specifically, the quality of object detection is directly related to the sample. When the sample contains only a single individual or the target is relatively prominent, the accuracy of object detection and individual recognition is high. However, the choice of target region detection and feature extraction methods also directly affects the final detection results. Furthermore, external environmental factors (light intensity, occlusion, etc.), shooting angle, and image quality all influence the detection results. Therefore, the method of constructing a multi-angle sample set to achieve object detection and image recognition in this invention still requires further exploration. Therefore, this invention aims to construct more complex sample sets (multi-angle, day and night) to simulate actual breeding scenarios, build intelligent algorithms applicable to different scenarios, and further develop an efficient, accurate, and easy-to-operate detection and recognition system.

[0032] Combination Figure 1 This embodiment describes a method for predicting important growth traits in pigs based on multi-source fusion, comprising:

[0033] S1. Data Acquisition and Processing:

[0034] Depth image acquisition and 3D point cloud conversion:

[0035] First, images of pigs were acquired using the Azure Kinect DK camera developed by Microsoft. This camera includes a depth sensor, an RGB sensor, a microphone array, and an infrared transmitter. The camera's depth sensor uses Time-of-Flight (TOF) measurement. The device is equipped with a 1-megapixel TOF imaging chip with advanced pixel technology, achieving an error within 2.0 mm when the measurement distance is less than 8.5 m.

[0036] Furthermore, the Azure Kinect DK camera offers multiple working modes and multi-device synchronization capabilities to meet the needs of various acquisition environments. Real-time monitoring of pig image acquisition quality is crucial for data collection during the acquisition process. This invention, based on the Microsoft SDK, utilizes Python to develop an automatic image acquisition program. This program can display acquired images in real time, named according to the timestamp of the capture, i.e., "year-month-day-hour-minute-second-microsecond". The OpenCV library is used to convert depth images into pseudo-color images in real time for image quality monitoring.

[0037] During depth image acquisition, Excel was used to record the ear tag, start time, and end time of each pig's acquisition, facilitating subsequent differentiation between pig images. Python was used to segment the image data for each pig, after which images with poses that did not meet standards or with partially occluded torsos were removed. A segmentation framework was employed for image segmentation.

[0038] Converting a depth image to a 3D point cloud is essentially a process of inverse projection from the pixel coordinate system to the camera coordinate system, which can be achieved simply by performing coordinate transformation using an intrinsic parameter matrix. During camera acquisition, point cloud data inevitably contains noise to varying degrees due to environmental interference, sensor errors, and the reflective properties of objects. Currently, many mature point cloud filtering algorithms exist, such as statistical filtering, pass-through filtering, radius filtering, and Gaussian filtering. Pig body surface point clouds are continuous, high-density data, rather than isolated, discrete point cloud data, making radius filtering a suitable method for noise reduction. This method effectively removes isolated discrete points and smooths the point cloud data, while also offering the advantage of easily adjustable parameters.

[0039] Based on the above steps, a unit for acquiring depth image data and processing 3D point clouds of the back and abdomen of pigs was developed using the Python language. It should be noted that this invention uses depth images to obtain 3D point clouds, which saves hardware costs. If hardware conditions permit, other methods can be used to obtain point cloud data.

[0040] Acceleration data acquisition: Acceleration data is acquired by using accelerometers installed on the pigs, such as... Figure 2 As shown, the combined acceleration of the three axes is expressed and recorded.

[0041] S2, Point Cloud Registration:

[0042] Point cloud registration technology unifies point clouds from two perspectives into the same coordinate system to complete the 3D reconstruction of a pig. The essence of point cloud registration is to find a suitable rigid transformation, that is, a rotation matrix and a translation vector, to transform the coordinate system of one set of point clouds to the coordinate system of another set of point clouds, thereby achieving the unification of the two coordinate systems.

[0043] This invention first performs coarse registration using the RANSAC algorithm or the Fast Global Registration (FRG) algorithm based on Fast Point Feature Histogram (FPFH) features. This roughly aligns the point cloud to an initial pose that facilitates rapid convergence of the subsequent fine registration algorithm. Subsequently, fine registration is performed using the Iterative Closest Point (ICP) algorithm to reduce the risk of the registration result deviating from the global optimum due to an undesirable initial pose.

[0044] The quality and efficiency of the ICP algorithm, the RANSAC coarse registration + ICP fine registration algorithm based on fast point feature histogram features (FPFH_RANSAC-ICP), and the FRG coarse registration + ICP fine registration algorithm (FRG-ICP) for pig 3D reconstruction were evaluated by calculating the interior point RMSE, single registration time, and the number of times the algorithm needs to be repeated to complete registration. This study explores and verifies the application effect of point cloud registration technology in pig 3D reconstruction, thereby constructing a complete surface of the pig and providing reliable technical support for further improving the accuracy of backfat thickness and eye muscle area prediction.

[0045] S3. Obtain body size parameter characteristics:

[0046] Research has revealed that backfat thickness and eye muscle area in pigs are related not only to the pig's growth stage but also to its morphology and body size. Therefore, this implementation method first uses LSSA_CAU software to measure body size. After automatically correcting the posture of the input point cloud data, key feature points are detected to measure the pig's body length, height, and width. The body length, height, and width are then normalized to obtain normalized body size parameter features. .

[0047] The body length, height, and width are measured as follows: When the pig is standing normally, the straight line length from the middle of the two ears to the base of the tail is the body length; the height from the top of the scapula to the ground is the body height; and the width of the pig is measured at the top of the scapula.

[0048] Subsequently, the datasets for backfat thickness and eye muscle area were divided into training and testing sets in an 8:2 ratio, with the data partitioning implemented using Python's sklearn library. This invention uses body length, body width, and body height as independent variables to predict backfat thickness and eye muscle area, constructing a predictive model.

[0049] It should be noted that the backfat thickness and eye muscle area data used for training can be obtained by using ultrasound equipment on live animals after slaughter. They are not needed for actual prediction and can be predicted by features such as body size.

[0050] S4. Morphological Feature Extraction:

[0051] There is a very close relationship between the thickness of backfat and the area and shape of the eye muscles in pigs. The shape can reflect the relevant data of backfat thickness and eye muscle area to a certain extent. However, this relationship cannot be directly represented by an expressible function. Therefore, this invention uses three-dimensional point cloud processing to reflect the relationship of shape.

[0052] Downsampling is performed on 3D point clouds to reduce the number of points in the cloud, thereby improving data processing efficiency, reducing computational costs, and preserving as much of the original point cloud's shape and features as possible. This invention uses farthest-per-second (FPS) sampling to downsample the point cloud data. It should be noted that this embodiment downsamples to 1024 points to accommodate computational power and efficiency. If the device's computing power can support the computational requirements, or if time constraints are not high, downsampling can be omitted, or the sampling rate can be set higher.

[0053] For the downsampled point cloud data, the Z-axis coordinates of the back point cloud data are converted into the pig's body height. The converted point cloud is then fed into a point cloud classification model with the classification layer removed, such as a PointNet or PointNet++ classification network with the last layer's sigmoid activation function removed. The model output is then used as the morphological feature. ;

[0054] It is important to note that while point clouds themselves represent morphological features, they cannot be directly used as features. The features represented by the point clouds need to be extracted. During this extraction process, since the output layer determines the spatial representation of the point cloud, this invention, after research, ultimately decided to use backfat thickness or eye muscle area as the output to guide the network model in feature extraction. Specifically, during PointNet or PointNet++ training, the output of the classification network, excluding the sigmoid activation function of the last layer, is set to backfat thickness or eye muscle area to guide the model's training process. In other words, the features represented in point cloud data form are transformed into a space that can represent backfat thickness or eye muscle area. Although the output features may appear to be backfat thickness or eye muscle area, they should essentially be understood as "morphological features" that can be represented in the backfat thickness or eye muscle area space.

[0055] It should also be noted that PointNet or PointNet++ is not trained together with the subsequent network models. Instead, PointNet or PointNet++ is trained separately. After training, the resulting features are actually spatial morphological features that can represent backfat thickness or eye muscle area. These morphological features are then fused together to participate in the training and prediction of subsequent models.

[0056] S5. Conduct correlation analysis to determine the indicators affecting back fat thickness or eye muscle area:

[0057] Backfat thickness and eye muscle area are closely related to pig breed type and other indicators. This invention studies multiple indicators. Since pig breed type needs to be determined in conjunction with genes, and considering that extracting gene data for subsequent actual prediction would be cumbersome, and also considering that pig breed is essentially a result of gene expression, the "cognized" pig breed type is directly used as a predictive indicator. Furthermore, since correlation analysis combining pig breed type with other factors can lead to uncertain correlations for other indicators, a holistic correlation analysis of pig breed type with other indicators is not performed. Instead, correlation analysis of other indicators is conducted separately for each pig breed type, and then other indicators are selected based on the correlation analysis results of different pig breed types. In the correlation analysis of pig growth indicators with backfat thickness and intramuscular fat content, phenotypic data of major growth traits of pigs such as Min pigs and Landrace pigs are used for analysis. The research group uses the inflection points of the gradual growth phase, rapid growth phase, and slow growth phase of the pig growth curve to divide the pig's growth trajectory into four stages: slow growth phase, rapid growth phase, deceleration growth phase, and plateau phase. Considering the convenience of later prediction, the study ultimately determined slaughter age and growth stage as the key influencing indicators. Correlation analysis of slaughter age, growth stage, backfat thickness, and IMF content in Min pigs revealed a significant positive correlation between slaughter age and backfat thickness. Compared to the deceleration growth stage, backfat thickness significantly increased during the plateau stage. The correlation coefficients between slaughter age, growth stage, and backfat thickness and IMF content are shown in Table 1.

[0058] Table 1. Correlation coefficients between slaughter age and growth stage and backfat thickness and IMF content.

[0059]

[0060] Note: p < 0.05 indicates a significant correlation; p < 0.01 indicates a highly significant correlation.

[0061] Furthermore, research has found a correlation between pig activity levels and backfat thickness or eye muscle area. However, there is currently no way to measure pig activity levels, nor is there a way to demonstrate the specific correlation between activity levels and backfat thickness or eye muscle area. Therefore, it is difficult to use pig activity levels for predicting backfat thickness or eye muscle area. This invention proposes a standard for measuring pig activity levels:

[0062] based on Figure 2Since data on the triaxial acceleration values ​​corresponding to the pig's movements are generally available, the amount of movement cannot be directly determined from the acceleration data. Therefore, this implementation method first collects acceleration data over a period of time (5 days in this implementation method to avoid the influence of random factors; in practice, other time lengths such as 1 day can also be chosen, depending on the actual situation). Figure 3 As shown, the median value of the statistical acceleration is denoted as threshold A. Then, for the daily acceleration data, using 10-minute intervals (10 minutes balances data volume and accuracy, but any duration, such as 5 minutes, can also be used), we count whether there are two or more acceleration data points greater than threshold A within a single unit. If so, the current unit is considered valid motion data; otherwise, it is not. We count the number N of valid motion data points for the day and normalize them to obtain the normalized daily valid motion volume. Here, 'n' represents n units within an hour. In this implementation, 10 minutes is considered one unit, therefore n=6. Finally, the average daily effective exercise volume within the slaughter age period is calculated. As a standard for the amount of exercise pigs do.

[0063] This implementation ultimately uses pig breed type, slaughter age, growth stage, and average daily effective exercise as indicators affecting backfat thickness or eye muscle area. It should be noted that, to facilitate processing by the neural network model, one-hot encoding is used for pig breed type and growth stage. Pig breed type is encoded using 4-bit binary code, and growth stage is encoded using 4-bit binary code. Each bit is then used as an element to construct a feature vector. The pig breed type feature vector, growth stage feature vector, slaughter age, and average daily effective exercise are concatenated to obtain the first feature vector. It should also be noted that slaughter age needs to be normalized, using the longest slaughter age as the benchmark.

[0064] It should also be noted that the 4-bit binary code for pig breed types in this embodiment is due to the sample size of the pig breed types used in this embodiment. The data used in this embodiment was collected from the breed registration information, growth and reproductive performance records of Landrace and Duroc pigs from a national-level core breeding farm. In the pedigree records, there were 8638 Landrace pigs born in 2017-2018, including 4420 boars and 4216 sows; and 4513 Duroc pigs, including 2289 boars and 2224 sows, which were divided into training and testing samples at an 8:2 ratio. These two breeds are divided into four breed types according to male and female, hence the use of a 4-bit binary code. When there are more pig breed types, a longer binary code can be used, but the corresponding network model will need to be retrained according to the scheme of this embodiment.

[0065] S6. Construction and training of neural network models:

[0066] In relevant studies (breeding value estimation), such as Figure 4 As shown, the BayesB method exhibits high accuracy in estimating breeding values ​​across all scenarios. However, in the practical field of predicting backfat thickness or eye muscle area, BayesB requires a significant amount of prior knowledge. This prior knowledge needs to be constructed, and since many factors influence backfat thickness or eye muscle area, different perspectives can lead to varying understandings of the prior knowledge. Therefore, the lack of prior knowledge may result in the BayesB network model not achieving optimal performance, and prediction stability may also be affected. Considering the convenience of subsequent predictions, this invention prioritizes neural network models as the primary prediction model. Furthermore, considering the performance of ANN models in related studies, this implementation adopts a CNN network model as the prediction model.

[0067] Furthermore, based on research findings, the representational ability of morphological features needs further improvement when classifying the aforementioned different types of feature data. Therefore, this invention ultimately decided not to directly concatenate body size parameters and morphological features to expand the data space and enhance spatial separability, but instead to use body size parameters to enhance morphological features: Then, the enhanced morphological features are fed into the Flatten layer to flatten them, resulting in the second feature vector.

[0068] The first and second feature vectors are concatenated and then used to make predictions through a simple CNN network model. The predicted output is... , The corresponding index values ​​for back fat thickness and eye muscle area.

[0069] Using pig breed type, slaughter age, growth stage, average daily effective exercise volume, and point cloud data as samples, and corresponding backfat thickness or eye muscle area as labels, a CNN network model is trained to obtain a well-trained CNN network model.

[0070] It should also be noted that although this embodiment uses a relatively simple CNN network model, the implementation process of the present invention is not limited to this relatively simple CNN network model structure, nor is it limited to the CNN network model. Other neural network models with stronger feature extraction and classification capabilities can also be used.

[0071] S7. Prediction by Neural Network Models:

[0072] The input features for the neural network model are obtained using steps S1-S5, and then the trained network model is used to predict the backfat thickness or eye muscle area. When conducting tests, since backfat thickness or eye muscle area can deviate from the slaughter measurement, this invention statistically calculates the error range for 1728 Landrace pigs (884 boars and 844 sows) and 903 Duroc pigs (458 boars and 445 sows). The error for backfat thickness is 12%-16%, and the error for eye muscle area is 11%-20%, which is already a very good prediction result.

[0073] The above examples of the present invention are merely illustrative of the computational model and process of the present invention, and are not intended to limit the implementation of the present invention. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is impossible to exhaustively list all possible implementations here. Any obvious variations or modifications derived from the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A method for predicting important growth traits of live pigs based on multi-source fusion, characterized in that, The method comprises the following steps: The body size parameters and morphological features of the live pig are obtained based on the corresponding point cloud data of the live pig, and then the body size parameter features are used to enhance the morphological features to obtain enhanced morphological features, which are flattened in a Flatten layer to obtain a second feature vector; a first feature vector is obtained based on the corresponding pig breed type feature vector, growth stage feature vector, and slaughter age and exercise amount of the live pig; The first feature vector and the second feature vector are spliced and then input into a neural network model to obtain a predicted output important growth trait prediction value.

2. The multi-source fusion-based pig important growth trait prediction method according to claim 1, characterized in that, The important growth trait prediction value includes the prediction values of backfat thickness and eye muscle area.

3. The multi-source fusion-based pig important growth trait prediction method according to claim 2, characterized in that, Before obtaining the body size parameters and morphological features of the live pig based on the corresponding point cloud data of the live pig, at least two views of point cloud data are obtained first, and point cloud registration is performed; after the point cloud registration, the point cloud data are used to obtain the body size parameters and morphological features of the live pig.

4. The multi-source fusion-based pig important growth trait prediction method according to claim 3, characterized in that, The process of point cloud registration comprises the following steps: First, coarse registration is performed based on a RANSAC algorithm or a fast global registration algorithm based on a fast point feature histogram feature, and then iterative closest point algorithm (ICP) is used for fine registration.

5. The multi-source fusion-based pig important growth trait prediction method according to claim 1, characterized in that, The process of obtaining the morphological features of the live pig based on the point cloud data adopts a PointNet or PointNet++ model with a sigmod activation function removed from the last layer.

6. The multi-source fusion based method for predicting important growth traits of pigs according to any one of claims 1 to 5, characterized in that, An enhanced morphological feature is obtained by enhancing the morphological feature using the body size parameter feature , is the transpose of , is a vector obtained by normalizing the body size parameter, i.e., the normalized body size parameter feature; is a morphological feature obtained based on the point cloud data.

7. The multi-source fusion-based method for predicting important growth traits of pigs according to claim 6, characterized in that, The exercise amount data used to obtain the first feature vector adopts average daily effective exercise amount, which is obtained by the following steps: Obtaining acceleration data of pigs in a period of time, counting the value corresponding to the median of the acceleration, denoted as threshold A; then for the acceleration data of each day, taking the acceleration data corresponding to the time length l as a unit, counting whether there are more than two acceleration data greater than the threshold A in a unit, if yes, denoted as the current unit as valid motion data, otherwise not as valid motion data; counting the number of valid motion data N in a day, and normalizing to obtain the normalized single-day valid motion amount Here n represents n units in an hour; Finally, the average daily effective movement amount within the slaughter age limit as the movement amount of the pig.

8. The multi-source fusion-based method for predicting important growth traits of pigs according to claim 7, characterized in that, During the process of obtaining the acceleration data of the pig in a period of time, 5 days are taken in the period of time; the time length l is 10 minutes.

9. The multi-source fusion-based method for predicting important growth traits of pigs according to claim 6, characterized in that, The pig breed type feature vector used to obtain the first feature vector is obtained by the following steps: Based on the pig breed type, onehot encoding is used for encoding, and then each bit of the encoding is taken as an element to form a feature vector.

10. The multi-source fusion-based method for predicting important growth traits of pigs according to claim 6, characterized in that, The growth stage feature vector used to obtain the first feature vector is obtained by the following steps: Based on the growth stage of the pig, onehot encoding is used for encoding, and then each bit of the encoding is taken as an element to form a feature vector.

Images