Livestock and poultry three-dimensional imaging measurement seed selection method and system based on millimeter wave radar

Abstract

The invention relates to the technical field of millimeter-wave radar and machine vision, and discloses a livestock and poultry three-dimensional imaging measurement seed selection method and system based on millimeter-wave radar, and the method comprises the steps: obtaining an electromagnetic scattering signal of livestock and poultry through the millimeter-wave radar; the laser radar is used for acquiring depth information of livestock and poultry; the camera is used for acquiring a video sequence of livestock and poultry; the signal processing module is used for recovering millimeter wave signals; the three-dimensional imaging module is used for converting the signal into a three-dimensional image; the motion compensation module is used for fusing the three-dimensional image, the depth information and the video sequence; the image processing module is used for extracting livestock and poultry body size parameters; the livestock and poultry seed selection module is used for screening seed sources based on livestock and poultry body size parameters; according to the invention, a millimeter wave radar imaging technology is adopted to solve the problem of non-contact three-dimensional shape measurement through animal fur.

Description

Technical Field

[0001] The present invention relates to the field of millimeter wave radar and machine vision technology, and more specifically, to a method and system for livestock and poultry three-dimensional imaging measurement and selection based on millimeter wave radar. Background Art

[0002] In the livestock and poultry industry, large-scale breeding is based on the measurement of the physical signs and dimensions of hundreds or thousands of farm animals. In order to better select and breed livestock and poultry, it is necessary to accurately measure the physiological structure dimensions of animals, such as bones and muscles. Livestock and poultry are easily frightened by workers during the inspection process, resulting in stress reactions, and the measurement efficiency and accuracy are low. Summary of the invention

[0003] The present invention provides a method and system for livestock and poultry three-dimensional imaging measurement and selection based on millimeter wave radar. The millimeter wave radar has the advantages of no harm to organisms, certain penetration of detection technology, and can penetrate livestock and poultry feathers, and low cost. The present invention uses millimeter wave radar to perform non-contact three-dimensional imaging of livestock and poultry, accurately measure the physical signs and body measurement parameters of livestock and poultry, and construct an efficient and accurate selection model. The invention will significantly improve the breeding efficiency and seed quality of livestock and poultry, thereby improving the production performance of the livestock and poultry industry and solving technical problems in related technologies.

[0004] The present invention provides a millimeter wave radar-based livestock and poultry three-dimensional imaging measurement and selection system, comprising: a millimeter wave radar for acquiring electromagnetic scattering signals of livestock and poultry;

[0005] LiDAR, to obtain depth information of livestock and poultry;

[0006] Camera, to obtain video sequences of livestock and poultry;

[0007] Also includes:

[0008] Signal processing module for millimeter wave signal recovery;

[0009] A three-dimensional imaging module, used for converting signals into three-dimensional images;

[0010] Motion compensation module, used to fuse 3D images, depth information and video sequences;

[0011] Image processing module, used to extract body size parameters of livestock and poultry;

[0012] The livestock and poultry selection module selects seed sources based on livestock and poultry body size parameters.

[0013] Furthermore, a rectangular channel is designed with a size just large enough for an animal to be tested to pass through, and millimeter-wave radar, lidar and camera are installed at different angles in the channel.

[0014] Furthermore, the antenna array design of the laser radar includes:

[0015] First, the antenna array is abstracted as a sparse array model. Assume that there are N antenna units in the array, the position of the nth antenna unit is (xn,yn,zn), and the position matrix of the array is defined as:

[0016] P = [x1, ..., x N ; y1,…,y N ; z1,…,z N ]

[0017] Each column represents the position coordinates of an antenna unit;

[0018] Assume that the imaging area is divided into M patches, and the scattering coefficient of each patch is σ m , then the observed signal is expressed as:

[0019]

[0020] Where a(θ m ,φ m ) is the steering vector corresponding to the mth patch, θ m and φ m are azimuth and elevation angles respectively, n is the noise;

[0021] The array position matrix P is optimized to minimize the error between the observed signal and the patch scattering coefficient:

[0022]

[0023] Where a(θ m ,φ m ; P) represents the relationship between the steering vector and the array position matrix;

[0024] The scattering coefficients are spatially aggregated using a patch-based dictionary, assuming that the patch dictionary is D = [d1, ..., d K ], where d k is the kth dictionary atom, then the scattering coefficient is expressed as a sparse representation of the dictionary:

[0025] σ=Dα

[0026] Where α is the sparse coefficient, and the dictionary representation is introduced into the array optimization model:

[0027]

[0028] Where λ is the regularization parameter, which is used to balance the data fitting term and the sparsity constraint term;

[0029] The patch size is optimized according to the estimation theory criterion, minimizing the Cramér-Rao lower bound. Let the patch size be Δ p , then CRLB is expressed as:

[0030]

[0031] where a n (θ m ,φ m ; P) is the steering vector of the nth antenna unit to the mth patch, is the noise variance, and the optimal patch size is obtained by minimizing CRLB:

[0032]

[0033] Combining the above models and constraints, the radar antenna array optimization model that meets the grating lobe suppression constraint is obtained:

[0034]

[0035] Where μ is the weight coefficient for patch size optimization. The optimization problem is solved by alternating optimization, that is, alternatingly updating P, α, and Δ p , until convergence.

[0036] A method for selecting livestock and poultry by three-dimensional imaging measurement based on millimeter wave radar comprises the following steps:

[0037] Step 101, the millimeter wave radar transmits a millimeter wave signal to illuminate livestock and poultry, and receives a scattered echo signal s(t, θ);

[0038] Step 102, performing compressed sensing reconstruction on the echo signal to obtain a scattering coefficient σ(x, y, z);

[0039] sts=Aσ

[0040] Where A is the observation matrix;

[0041] Step 103: Use the back projection algorithm to perform three-dimensional imaging on the scattering coefficient to obtain a millimeter wave image I mmw (x, y, z);

[0042] Step 104, using photoelectric fusion technology to compensate for livestock and poultry motion, and obtain a compensated image I motion ;

[0043]

[0044] in To compensate the operator, it is implemented through methods such as optical flow;

[0045] Step 105, using a deep learning algorithm from I motion Extracting livestock and poultry body size parameters {l i};

[0046]

[0047] in is a deep network, Θ is a network parameter;

[0048] Step 106: according to the extracted body size parameters {l i}Carry out livestock and poultry breeding.

[0049] Furthermore, by transmitting millimeter wave signals to illuminate livestock and poultry targets and receiving scattered echo signals of the targets, assuming that the transmitted signal is s(t, θ) and the scattering coefficient of the target is σ(x, y, z), the received signal r(t, θ) is expressed as:

[0050]

[0051] in is the distance between the radar and the target, c is the speed of light, (x r ,y r , z r ) are the coordinates of the radar.

[0052] Furthermore, the signal is reconstructed using compressed sensing theory, and the observed signal is discretized and expressed in matrix form:

[0053] R=A∑+N

[0054] Where R is the observation signal matrix, A is the observation matrix, ∑ is the scattering coefficient matrix, and N is the noise matrix. The goal of compressed sensing is to recover the sparse ∑ from R, which is achieved by solving the following l1 norm minimization problem:

[0055] st||RA∑||2≤ε

[0056] Where ε is the noise tolerance, solving the optimization problem to obtain the reconstructed scattering coefficient matrix

[0057] Furthermore, the reconstructed scattering coefficient is used to realize three-dimensional imaging through the back projection algorithm. Assuming the coordinates of the imaging area are (x, y, z), the back projection imaging process is expressed as:

[0058]

[0059] Where I(x, y, z) is a three-dimensional image, is the reconstructed scattering coefficient, θ iis the i-th observation angle, N θ is the total number of observation angles.

[0060] Furthermore, a deep learning algorithm is used to automatically extract the body size parameters of livestock and poultry from the compensated 3D images, and an end-to-end convolutional neural network is designed. Its input is I motion , the output is the body size parameter {l i}:

[0061]

[0062] Where Θ is the network parameter, which is trained by supervised learning:

[0063]

[0064] in is the training data set.

[0065] Furthermore, a comprehensive evaluation function is set Weighted sum of different body size parameters:

[0066]

[0067] where w i is the weight coefficient of the i-th body size parameter, N l is the total number of body size parameters, sort the evaluation function values, and select the top S individuals as the provenance.

[0068] A computer storage medium stores non-temporary computer-readable instructions capable of executing one or more steps in the aforementioned livestock and poultry three-dimensional imaging measurement and selection method based on millimeter wave radar.

[0069] The beneficial effects of the present invention are:

[0070] The present invention uses millimeter wave radar imaging technology to solve the problem of non-contact three-dimensional shape measurement through animal fur. There are currently no similar research reports in related fields. The present invention is expected to fill the gap in this field and empower the livestock and poultry farming industry to judge the development of raised animals and select breeding technologies.

[0071] The present invention conducts research on millimeter-wave radar array design suitable for three-dimensional imaging of animal body features, explores sparse MIMO array optimization design methods, and creates a new scenario for millimeter-wave radar applications.

[0072] The present invention develops a CS algorithm to perform rapid target image reconstruction, providing a feasible back-end image reconstruction processing technology path for achieving real-time three-dimensional imaging of livestock and poultry bodies.

[0073] The present invention is an ideal application mode for directly imaging animals in a walking state. Based on the micro-motion model of walking animals, the present invention proposes to use photoelectric recognition technology to extract animal motion parameters as prior knowledge of motion imaging algorithms, and applies the electromagnetic scattering mechanism of the micro-motion model of animals in a walking state and the optimal layout theory of reconfigurable sparse arrays to animal three-dimensional imaging, so as to solve the problem of animal motion imaging and improve detection efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0074] Figure 1 This is a diagram of the components of a millimeter wave radar-based livestock and poultry three-dimensional imaging measurement and selection system of the present invention;

[0075] Figure 2 It is a schematic diagram of the experimental scene of the present invention;

[0076] Figure 3 It is the display and control interface of the livestock and poultry multi-sensory fusion non-contact imaging measurement system of the present invention. DETAILED DESCRIPTION

[0077] The subject matter described herein will now be discussed with reference to example implementations. It should be understood that the discussion of these implementations is only to enable those skilled in the art to better understand and implement the subject matter described herein, and the functions and arrangements of the elements discussed may be changed without departing from the scope of protection of the present specification. Various examples may omit, replace, or add various processes or components as needed. In addition, the features described in some examples may also be combined in other examples.

[0078] The design of a livestock and poultry three-dimensional imaging measurement and selection system based on millimeter-wave radar needs to overcome the following problems:

[0079] 1. Radar antenna array design problem that meets grating lobe suppression constraints

[0080] Antenna array design is one of the key technologies in building imaging systems and is closely related to the study of imaging algorithms. The formal design of the antenna array is directly related to the subsequent imaging algorithm and is therefore crucial for the optimization of the radar antenna array design. The present invention first uses a sparse unit array to establish a patch-based sparse array optimization model, which is very different from classical radar, which mainly focuses on long-range settings where targets are well modeled as point scatterers. For short-range sparse arrays, point scatterer target models lead to grating lobes. It is proposed to study a patch-based target model that is appropriately optimized for sensor and scene geometry to suppress such grating lobes. The grating lobes are suppressed by spatial aggregation through a patch-based dictionary, and the patch size can be optimized according to estimation theory criteria.

[0081] 2. Design of radar signal fast imaging algorithm with real-time computational complexity constraints

[0082] For millimeter-wave animal 3D imaging, the size of the millimeter-wave image of animal targets is in the order of hundreds of thousands or even millions of pixels, which is too large for ordinary computers to store and process. Compressed sensing (CS) is a signal processing technology that can compress data below the Nyquist rate. This technology can use the sparsity of the signal to recover the signal from much fewer samples than required by the Nyquist-Shannon sampling theory. This advantage is very effective in system applications with large amounts of data. A convex optimization model for compressed sensing multiple-input multiple-output (MIMO) array design is proposed. It is proposed to study the issues related to focusing, sidelobe suppression, and grating lobe suppression of the synthetic array and the corresponding fast imaging algorithms.

[0083] 3. Design of imaging compensation algorithm for non-cooperative moving target detection

[0084] Since the imaging distance is near field, the back projection (BP) algorithm is generally used to achieve imaging. The principle of the BP algorithm is to perform phase compensation for each point in the imaging area and coherently accumulate the compensated echo data, so the calculation time is relatively long. Therefore, it is impossible to perform real-time imaging of animals in a walking state, and improvements must be made at the algorithm level to significantly reduce the amount of calculation. It is proposed to use photoelectric recognition technology to extract animal motion parameters as prior knowledge of motion imaging algorithms, and apply the electromagnetic scattering mechanism of animal micro-motion models in a walking state and the optimal array theory of reconfigurable sparse arrays to animal three-dimensional imaging.

[0085] The system integrates millimeter-wave radar, lidar, camera and other sensors to perform non-contact imaging of the geese and obtain their body size.

[0086] At least one embodiment of the present invention discloses a livestock and poultry three-dimensional imaging measurement and selection system based on millimeter wave radar, such as Figure 1 As shown, including:

[0087] The millimeter wave radar obtains the electromagnetic scattering signal s(t, θ) of livestock and poultry, where t is time and θ is the observation angle.

[0088] The laser radar obtains the depth information d(x, y) of livestock and poultry, where x and y are plane coordinates.

[0089] The camera captures video sequences of livestock and poultry Among them I n is the nth frame image;

[0090] MIMO signal processing module based on compressed sensing for millimeter wave signal recovery.

[0091] The three-dimensional imaging module based on back projection is used to convert the signal into a three-dimensional image.

[0092] The motion compensation module based on optoelectronic fusion is used to fuse three-dimensional images, depth information and video sequences to compensate for imaging blur caused by livestock and poultry movement.

[0093] A deep learning-based image processing module for extracting livestock and poultry body size parameters.

[0094] The livestock and poultry selection module selects seed sources based on livestock and poultry body size parameters.

[0095] At least one embodiment of the present invention discloses a method for selecting livestock and poultry by three-dimensional imaging measurement based on millimeter wave radar, comprising the following steps:

[0096] Step 101, the millimeter wave radar transmits a millimeter wave signal to illuminate livestock and poultry, and receives a scattered echo signal s(t, θ);

[0097] Step 102, performing compressed sensing reconstruction on the echo signal to obtain a scattering coefficient σ(x, y, z);

[0098] sts=Aσ

[0099] Where A is the observation matrix.

[0100] Step 103: Use the back projection algorithm to perform three-dimensional imaging on the scattering coefficient to obtain a millimeter wave image I mmw (x, y, z);

[0101]

[0102] Step 104, using photoelectric fusion technology to compensate for livestock and poultry motion, and obtain a compensated image I motion ;

[0103]

[0104] in To compensate the operator, it can be achieved through methods such as optical flow;

[0105] Step 105, using a deep learning algorithm from I motion Extracting livestock and poultry body size parameters {l i};

[0106]

[0107] in is a deep network, Θ is a network parameter;

[0108] Step 106: according to the extracted body size parameters {l i} Carry out livestock and poultry breeding;

[0109] The following is an explanation of the specific implementation of the aforementioned components or steps:

[0110] By transmitting millimeter wave signals to illuminate livestock and poultry targets, the scattered echo signals of the targets are received. Assuming that the transmitted signal is s(t, θ) and the scattering coefficient of the target is σ(x, y, z), the received signal r(t, θ) can be expressed as:

[0111]

[0112] in is the distance between the radar and the target, c is the speed of light, (x r ,y r , z r ) are the coordinates of the radar;

[0113] Since the radar observation data is incomplete, it is necessary to use the compressed sensing theory to reconstruct the signal. Discretize the observation signal and express it in matrix form:

[0114] R=A∑+N

[0115] Where R is the observed signal matrix, A is the observation matrix, ∑ is the scattering coefficient matrix, and N is the noise matrix. The goal of compressed sensing is to recover the sparse ∑ from R, which can be achieved by solving the following l1 norm minimization problem:

[0116] st||RA∑||2≤ε

[0117] Where ε is the noise tolerance. Solving this optimization problem can obtain the reconstructed scattering coefficient matrix

[0118] Backprojection 3D imaging

[0119] Using the reconstructed scattering coefficient, three-dimensional imaging can be achieved through the back projection algorithm. Assuming the coordinates of the imaging area are (x, y, z), the back projection imaging process can be expressed as:

[0120]

[0121] Where I(x, y, z) is a three-dimensional image, is the reconstructed scattering coefficient, θ i is the i-th observation angle, N θ is the total number of observation angles.

[0122] Optical-electric fusion motion compensation

[0123] Since livestock and poultry will have certain movements when walking, resulting in blurred images. In order to compensate for the motion blur, photoelectric fusion technology can be used to fuse the motion information of the optical image into the millimeter wave image.

[0124] Deep learning body size extraction

[0125] Using deep learning algorithms, the body size parameters of livestock and poultry can be automatically extracted from compensated three-dimensional images. Design an end-to-end convolutional neural network Its input is I motion , the output is the body size parameter {l i}:

[0126]

[0127] Where Θ is the network parameter, which can be trained by supervised learning:

[0128]

[0129] in is the training data set;

[0130] Livestock and poultry selection

[0131] According to the extracted body size parameters {l i}, combining knowledge from fields such as genetics and biology, designing livestock and poultry selection algorithms to screen out excellent seed sources.

[0132] For example, you can set a comprehensive evaluation function Weighted sum of different body size parameters:

[0133]

[0134] where w i is the weight coefficient of the i-th body size parameter, N l is the total number of body size parameters. By sorting the evaluation function values, the best livestock and poultry individuals can be selected.

[0135] Sparse Cell Array Modeling

[0136] First, the antenna array is abstracted as a sparse array model. Assume that there are N antenna units in the array, and the position of the nth antenna unit is (xn,yn,zn). The position matrix of the array is defined as:

[0137] P = [x1, ..., x N ; y1,…,y N ; z1,…,z N ]

[0138] Each column represents the location coordinates of an antenna unit.

[0139] Patch-based sparse array optimization model

[0140] The traditional point scatterer model will lead to grating lobes, so a patch-based target model is introduced. Assume that the imaging area is divided into M patches, and the scattering coefficient of each patch is σ m , then the observed signal can be expressed as:

[0141]

[0142] Where a(θ m ,φ m ) is the steering vector corresponding to the mth patch, θ m and φ m are the azimuth and elevation angles respectively, and n is the noise.

[0143] In order to suppress grating lobes, the array position matrix P needs to be optimized to minimize the error between the observed signal and the patch scattering coefficient:

[0144]

[0145] Where a(θ m ,φ m ; P) represents the relationship between the steering vector and the array position matrix.

[0146] Spatial aggregation based on patch dictionary

[0147] To further suppress grating lobes, the scattering coefficients can be spatially aggregated using a patch-based dictionary. Let the patch dictionary be D = [d1, ..., d K ], where d k is the kth dictionary atom, then the scattering coefficient can be expressed as a sparse representation of the dictionary:

[0148] σ=Dα

[0149] Where α is the sparse coefficient. The dictionary representation is introduced into the array optimization model:

[0150]

[0151] Where λ is the regularization parameter, which is used to balance the data fitting term and the sparsity constraint term.

[0152] Patch size optimization

[0153] The choice of patch size has an important impact on the grating lobe suppression performance. The patch size can be optimized according to estimation theory criteria, such as minimizing the Cramér-Rao lower bound (CRLB) or maximizing the mutual information (MI). Taking CRLB as an example, let the patch size be Δ p , then CRLB can be expressed as:

[0154]

[0155] where a n (θ m ,φ m ; P) is the steering vector of the nth antenna unit to the mth patch, is the noise variance. The optimal patch size can be obtained by minimizing CRLB:

[0156]

[0157] Array Optimization Algorithm

[0158] Combining the above models and constraints, we can obtain the radar antenna array optimization model that meets the grating lobe suppression constraint:

[0159]

[0160] Where μ is the weight coefficient for patch size optimization. This optimization problem can be solved by alternating optimization, that is, alternatingly updating P, α, and Δ p , until convergence.

[0161] Experimental scenario: Design a rectangular channel with a size that is just large enough for an animal to be tested to pass through. Figure 2 As shown in the figure, by installing millimeter wave radar, laser radar and other equipment at different angles in the channel, the animal can be fully covered. When the animal passes through the channel, the imaging system is started to perform three-dimensional imaging of the animal. The measurement of physical signs and body size is completed on the imaging interface, as shown in the figure. Figure 3 As shown. By collecting body size data and designing a selection and screening algorithm, the purpose of livestock and poultry selection can be achieved.

[0162] The present invention is committed to using advanced scientific and technological means such as visual recognition and radar measurement to accurately measure the body parameters of livestock and poultry, and to construct an efficient and accurate selection model. First, it will significantly improve the breeding efficiency and seed quality of livestock and poultry. Through precise breeding, better quality varieties can be cultivated, thereby improving production performance. Secondly, the application of advanced technology will greatly shorten the breeding cycle, accelerate genetic progress, and enable farmers to respond to market demand more quickly. Furthermore, the automated measurement and selection system will greatly reduce the cost of manual selection and improve labor productivity. These factors combined will significantly increase the benefits of breeding, enhance the overall competitiveness of the livestock and poultry industry, drive the coordinated development of upstream and downstream industrial chains such as related feed, slaughtering and processing, logistics and transportation, form a benign industrial ecosystem, and inject new vitality into local economic development.

[0163] The above describes an embodiment of the present invention, but this embodiment is not limited to the above-mentioned specific implementation mode. The above-mentioned specific implementation mode is merely illustrative and not restrictive. Under the guidance of this embodiment, ordinary technicians in this field can also make more forms of equivalent embodiments, all of which are within the protection of this embodiment.

Claims

1. A livestock and poultry three-dimensional imaging measurement and selection system based on millimeter wave radar, comprising: Millimeter wave radar, to obtain electromagnetic scattering signals from livestock and poultry; LiDAR, to obtain depth information of livestock and poultry; Camera, to obtain video sequences of livestock and poultry; It is characterized by further comprising: Signal processing module for millimeter wave signal recovery; A three-dimensional imaging module, used for converting signals into three-dimensional images; Motion compensation module, used to fuse 3D images, depth information and video sequences; Image processing module, used to extract body size parameters of livestock and poultry; The livestock and poultry selection module selects seed sources based on livestock and poultry body size parameters.

2. The livestock and poultry three-dimensional imaging measurement and selection system based on millimeter wave radar according to claim 1 is characterized in that: Design a rectangular channel with the size of an animal to be tested, and install millimeter-wave radar, lidar and camera at different angles in the channel.

3. The livestock and poultry three-dimensional imaging measurement and selection system based on millimeter wave radar according to claim 1 is characterized in that: The antenna array design of LiDAR includes: First, the antenna array is abstracted as a sparse array model. Assume that there are N antenna units in the array, the position of the nth antenna unit is (xn, yn, zn), and the position matrix of the array is defined as: P=[x1,...,x N (y1,...,y N !!z1,...,z N ] Each column represents the position coordinates of an antenna unit; Assume that the imaging area is divided into M patches, and the scattering coefficient of each patch is σ m , then the observed signal is expressed as: Where a(θ m ,φ m ) is the steering vector corresponding to the mth patch, θ m and φ m are azimuth and elevation angles respectively, n is the noise; The array position matrix P is optimized to minimize the error between the observed signal and the patch scattering coefficient: Where a(θ m ,φ m ; P) represents the relationship between the steering vector and the array position matrix; The scattering coefficients are spatially aggregated using a patch-based dictionary, assuming that the patch dictionary is D = [d1, ..., d K ], where d k is the kth dictionary atom, then the scattering coefficient is expressed as a sparse representation of the dictionary: σ=Dα Where α is the sparse coefficient, and the dictionary representation is introduced into the array optimization model: Where λ is the regularization parameter, which is used to balance the data fitting term and the sparsity constraint term; The patch size is optimized according to the estimation theory criteria to minimize the Cramér-Rao lower bound. Let the patch size be Δp, then CRLB is expressed as: where a n (θ m ,φ m ; P) is the steering vector of the nth antenna unit to the mth patch, is the noise variance, and the optimal patch size is obtained by minimizing CRLB: Combining the above models and constraints, the radar antenna array optimization model that meets the grating lobe suppression constraint is obtained: Where μ is the weight coefficient for patch size optimization. The optimization problem is solved by alternating optimization, that is, alternatingly updating P, α, and Δ p , until convergence.

4. A method for selecting livestock and poultry by three-dimensional imaging measurement based on millimeter wave radar, characterized in that: The following steps are involved: Step 101, the millimeter wave radar transmits a millimeter wave signal to illuminate livestock and poultry, and receives a scattered echo signal s(t, θ); Step 102, performing compressed sensing reconstruction on the echo signal to obtain a scattering coefficient σ(x, y, z); Step 103: Use the back projection algorithm to perform three-dimensional imaging on the scattering coefficient to obtain a millimeter wave image I mmw (x, y, z); Step 104, using photoelectric fusion technology to compensate for livestock and poultry motion, and obtain a compensated image I motion ; Its reason To compensate the operator, it is implemented through methods such as optical flow; Step 105, using a deep learning algorithm from I motion Extracting livestock and poultry body size parameters {l i }; Its reason is a deep network, Θ is a network parameter; Step 106: according to the extracted body size parameters {l i }Carry out livestock and poultry breeding.

5. The method for selecting livestock and poultry by three-dimensional imaging measurement based on millimeter wave radar according to claim 4, characterized in that: By transmitting millimeter wave signals to illuminate livestock and poultry targets, the scattered echo signals of the targets are received. Assuming that the transmitted signal is s(t, θ) and the scattering coefficient of the target is σ(x, y, z), the received signal r(t, θ) is expressed as: in is the distance between the radar and the target, c is the speed of light, (x r ,y r , z r ) are the coordinates of the radar.

6. The method for selecting livestock and poultry by three-dimensional imaging measurement based on millimeter wave radar according to claim 4 is characterized in that: The signal is reconstructed using compressed sensing theory, and the observed signal is discretized and expressed in matrix form: R=A∑+N Where R is the observation signal matrix, A is the observation matrix, ∑ is the scattering coefficient matrix, and N is the noise matrix. The goal of compressed sensing is to recover the sparse ∑ from R, which is achieved by solving the following l1 norm minimization problem: Where ε is the noise tolerance, solving the optimization problem to obtain the reconstructed scattering coefficient matrix 7. The method for selecting livestock and poultry by three-dimensional imaging measurement based on millimeter wave radar according to claim 4, characterized in that: Using the reconstructed scattering coefficient, three-dimensional imaging is achieved through the back projection algorithm. Assuming the coordinates of the imaging area are (x, y, z), the back projection imaging process is expressed as: Where I(x, y, z) is a three-dimensional image, is the reconstructed scattering coefficient, θ i is the i-th observation angle, N θ is the total number of observation angles.

8. According to claim 4, a method for selecting livestock and poultry by three-dimensional imaging measurement based on millimeter wave radar is characterized in that: Use deep learning algorithms to automatically extract the body size parameters of livestock and poultry from compensated 3D images and design an end-to-end convolutional neural network Its input is I motion , the output is the body size parameter {l i }: Where Θ is the network parameter, which is trained by supervised learning: Its reason is the training data set.

9. The method for selecting livestock and poultry by three-dimensional imaging measurement based on millimeter wave radar according to claim 4, characterized in that: Set up a comprehensive evaluation function Weighted sum of different body size parameters: where w i is the weight coefficient of the i-th body size parameter, N l is the total number of body size parameters, sort the evaluation function values, and select the top S individuals as the provenance.

10. A computer storage medium, characterized in that: Non-temporary computer-readable instructions are stored, and one or more steps in a livestock and poultry three-dimensional imaging measurement and selection method based on millimeter-wave radar as described in any one of claims 4-9 can be executed.

Images