A Lameness Recognition Method for Dairy Cows Based on Cow Body Posture Estimation

Abstract

This invention provides a method for identifying lameness in dairy cows based on cow posture estimation, comprising: acquiring a walking side view image of a dairy cow; constructing a dairy cow posture estimation model; inputting the walking side view image into the dairy cow posture estimation model to obtain feature vectors; and identifying the feature vectors to determine whether the dairy cow is lame. This invention improves the accuracy of lameness detection in dairy cows, which is of great significance for improving dairy cow welfare and reducing the economic losses caused by lameness, and has certain application prospects.

Description

Technical Field

[0001] This invention belongs to the field of posture estimation technology, and in particular relates to a method for recognizing lameness in dairy cows based on bovine posture estimation. Background Technology

[0002] Lameness is not only one of the most important welfare issues for dairy cows, but it also reduces milk production, impairs reproductive performance, and leads to significant economic losses. Prevention and timely treatment are the two main strategies for addressing lameness in dairy cows. Early identification of lame cows will reduce economic losses and aid in recovery. Therefore, identifying cows in the early stages of lameness is of great research significance.

[0003] Traditional methods for detecting lameness in dairy cows are manual, time-consuming, and labor-intensive, no longer meeting the needs of the current dairy farming industry. To address this issue, machine vision methods are being applied to lameness detection. A representative approach involves using pose estimation, which utilizes static images or video information to design specific feature descriptors (such as color, texture, shape, edges, and spatial features) to create feature vectors. These vectors are then used as input to a classifier trained using machine learning. The feature vectors to be identified are fed into this classifier for training, resulting in classification results for different animal pose categories. The pose feature descriptor is a crucial part of pose recognition, directly affecting the final recognition accuracy. Machine vision methods, by detecting lameness in dairy cows without interfering with their normal activities, offer advantages such as non-contact, stress-free operation, and low cost, and have been widely studied.

[0004] Existing machine vision-based methods for detecting lameness in dairy cows include calculating the curvature of the cow's back and fitting a straight line to the head and neck as lameness parameters; and analyzing the leg swing curves to extract lameness indicators for automatic detection. However, existing methods have low coverage of lameness feature extraction and cannot track key parts of the cow's body in complex environments to obtain more accurate lameness parameters. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention proposes a lameness identification method for dairy cows based on cow posture estimation. This method improves the accuracy of lameness detection and is of great significance for improving dairy cow welfare and reducing the economic losses caused by lameness, thus showing promising application prospects.

[0006] To achieve the above objectives, the present invention provides a method for identifying lameness in dairy cows based on cow posture estimation, comprising: acquiring a walking side view image of a dairy cow;

[0007] Construct a cow posture estimation model;

[0008] The walking side view image is input into the cow posture estimation model to obtain feature vectors;

[0009] The feature vector is identified to determine whether the cow has become lame.

[0010] Optionally, constructing the cow posture estimation model includes:

[0011] Construct a dataset; the dataset includes: side view images of cows walking normally and side view images of cows lame;

[0012] Label the dataset;

[0013] Construct a ResNet101 convolutional neural model;

[0014] Based on the labeled dataset, the ResNet101 convolutional neural model is trained to obtain the cow pose estimation model.

[0015] Optionally, labeling the dataset includes:

[0016] The cow's body and joints in the walking side view images of the cow in the dataset are marked; wherein, the joints include: the cow's mouth, forehead, neck, spine, tailbone, front left leg root, front left knee, front left hoof, front right leg root, front right knee, front right hoof, rear left leg root, rear left knee, rear left hoof, rear right leg root, rear right knee, and rear right hoof.

[0017] Optionally, the feature vector includes: a cow forelimb feature vector and a cow hindlimb feature vector.

[0018] Optionally, obtaining the feature vector includes:

[0019] The walking side view image is input into the cow posture estimation model to obtain the cow's skeleton sequence; the skeleton sequence includes: a normal walking skeleton sequence and a lame skeleton sequence;

[0020] Based on the skeletal sequence, the relative distance between the hooves of the cow is obtained; the hoof distance is the relative distance between the forelimbs and hindlimbs.

[0021] The feature vector is constructed based on the relative distance.

[0022] Optionally, obtaining the skeletal sequence of the cow includes:

[0023] The walking side view image is input into the cow posture estimation model to obtain the feature image;

[0024] Using non-maximum suppression, a cow target detection box is obtained based on the feature image;

[0025] The cow pose estimation model is used to identify the cow target detection box and obtain the cow's various joints and the connections between them.

[0026] The skeleton sequence is obtained based on the joints and the lines connecting the joints.

[0027] Optionally, the relative distance is:

[0028]

[0029]

[0030] Where x1, x2, y1, and y2 are the positions of the cow's left foreleg, right foreleg, left hindleg, and right hindleg, respectively, and L2 前 L2 后 These represent the relative distances between the cow's forelimbs and hindlimbs.

[0031] Optionally, the ResNet101 convolutional neural model includes: 3×3 convolutional blocks, 2×2 max pooling layers, and softmax layers.

[0032] Optionally, training the ResNet101 convolutional neural model includes:

[0033] The labeled dataset is input into the ResNet101 convolutional neural model to obtain the confidence map of each joint location;

[0034] The location of the joint is predicted based on the global maximum representative network in the confidence graph, thereby obtaining the location of the cow's body and the joint.

[0035] Compared with the prior art, the present invention has the following advantages and technical effects:

[0036] (1) Construct a cow pose estimation network model to extract the skeleton sequence of cow walking videos in natural living environment. The deep learning model algorithm has strong generalization ability and can effectively avoid false detection and missed detection caused by lighting, occlusion and angle problems.

[0037] (2) A cow posture estimation network is used to obtain the changing positions of the cow's body joints. Based on the walking change pattern of the cow's hoof joints, feature vectors of the cow's forelimbs and hindlimbs are constructed. The Bi-LSTM network, which is good at handling sequence problems, is used to distinguish whether the cow is lame. This not only reduces the computational complexity but also greatly improves the computational speed and effectively improves the recognition accuracy. Attached Figure Description

[0038] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings:

[0039] Figure 1 This is a schematic diagram of the method flow according to an embodiment of the present invention;

[0040] Figure 2 This is a schematic diagram of video capture of cows walking according to an embodiment of the present invention;

[0041] Figure 3 This is a schematic diagram of cow image pose recognition according to an embodiment of the present invention;

[0042] Figure 4 This is a schematic diagram of the results of cow side-view walking video recognition according to an embodiment of the present invention;

[0043] Figure 5 This is a schematic diagram of the relative distance change curve of the limbs in an embodiment of the present invention; wherein, (a) is a schematic diagram of the relative distance change curve of the forelimbs and hindlimbs of a dairy cow walking normally, and (b) is a schematic diagram of the relative distance change curve of the forelimbs and hindlimbs of a dairy cow lame.

[0044] Figure 6 This is a schematic diagram of limp recognition and classification according to an embodiment of the present invention;

[0045] Figure 7 This is a schematic diagram of the ResNet101 convolutional neural network according to an embodiment of the present invention. Detailed Implementation

[0046] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0047] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0048] This invention proposes a method for identifying lameness in dairy cows based on bovine posture estimation, comprising:

[0049] Obtain a walking side view image of a cow;

[0050] Construct a cow posture estimation model;

[0051] The walking side view image is input into the cow posture estimation model to obtain feature vectors;

[0052] The feature vector is identified to determine whether the cow has become lame.

[0053] Furthermore, constructing the cow posture estimation model includes:

[0054] Construct a dataset; the dataset includes: side view images of cows walking normally and side view images of cows lame;

[0055] Label the dataset;

[0056] Construct a ResNet101 convolutional neural model;

[0057] Based on the labeled dataset, the ResNet101 convolutional neural model is trained to obtain the cow pose estimation model.

[0058] Furthermore, labeling the dataset includes:

[0059] The cow's body and joints in the walking side view images of the cow in the dataset are marked; wherein, the joints include: the cow's mouth, forehead, neck, spine, tailbone, front left leg root, front left knee, front left hoof, front right leg root, front right knee, front right hoof, rear left leg root, rear left knee, rear left hoof, rear right leg root, rear right knee and rear right hoof.

[0060] Furthermore, the feature vectors include: cow forelimb feature vectors and cow hindlimb feature vectors.

[0061] Further, obtaining the feature vector includes:

[0062] The walking side view image is input into the cow posture estimation model to obtain the cow's skeleton sequence; the skeleton sequence includes: a normal walking skeleton sequence and a lame skeleton sequence;

[0063] Based on the skeletal sequence, the relative distance between the hooves of the cow is obtained; the hoof distance is the relative distance between the forelimbs and hindlimbs.

[0064] The feature vector is constructed based on the relative distance.

[0065] Furthermore, obtaining the skeletal sequence of the cow includes:

[0066] The walking side view image is input into the cow posture estimation model to obtain the feature image;

[0067] Using non-maximum suppression, a cow target detection box is obtained based on the feature image;

[0068] The cow pose estimation model is used to identify the cow target detection box and obtain the cow's various joints and the connections between them.

[0069] The skeleton sequence is obtained based on the joints and the lines connecting the joints.

[0070] Furthermore, the relative distance is:

[0071]

[0072]

[0073] Where x1, x2, y1, and y2 are the positions of the cow's left foreleg, right foreleg, left hindleg, and right hindleg, respectively, and L2 前 L2 后 These represent the relative distances between the cow's forelimbs and hindlimbs.

[0074] Furthermore, the ResNet101 convolutional neural model includes: 3×3 convolutional blocks, 2×2 max pooling layers, and softmax layers.

[0075] Furthermore, training the ResNet101 convolutional neural model includes:

[0076] The labeled dataset is input into the ResNet101 convolutional neural model to obtain the confidence map of each joint location;

[0077] The location of the joint is predicted based on the global maximum representative network in the confidence graph, thereby obtaining the location of the cow's body and the joint.

[0078] Example

[0079] The technical solution of this embodiment is: a method for recognizing lameness in dairy cows based on cow posture estimation, the steps of which include: 1) firstly, collecting videos of dairy cows walking in a farm environment; 2) constructing and training a dairy cow posture estimation network based on a ResNet101 convolutional neural network; 3) recognizing the posture of dairy cows based on the videos of dairy cow walking and saving the skeletal sequence of dairy cows and obtaining the feature vectors of the front and hind limbs of dairy cows; 4) recognizing whether dairy cows are lame based on the skeletal sequence of dairy cows and a Bi-LSTM classifier network.

[0080] In step 1):

[0081] 1.1) For step 1), use a camera fixed to the outside of the milking aisle to capture side-view walking videos of the cows;

[0082] 1.2) For step 1.1), the camera is about 2.5 meters away from the milking passage, and ensures that the cow can walk at least two movement cycles within one screen of the camera (one movement cycle is defined as the cow raising and lowering all four hooves once).

[0083] In step 2):

[0084] 2.1) For the cow walking videos collected in step 1), select 200 representative n-frame videos of cows walking normally and limping.

[0085] 2.2) For the cow walking videos selected in step 2.1), select 50 videos of cows walking normally and 50 videos of cows limping, and then select 20 frames from each video segment in sequence, for a total of 2000 cow walking side view images. The steps include:

[0086] 2.2.1) Use Labelme software to manually label 2000 side-view images of dairy cows walking, marking the cow's body and key points;

[0087] 2.2.2) Train model weights on the COCO dataset using transfer learning;

[0088] 2.2.3) A residual network ResNet101 is used to construct a cow pose estimation network to extract basic features of cow images and realize target detection and cow body joint detection in cow images;

[0089] 2.2.4) Use the `ploy` function in Python's matplotlib image library to generate images of cow pose estimation;

[0090] In step 3):

[0091] 3.1) Using the cow side-view walking video obtained in step 1.2) as input, construct a cow pose estimation network in step 2.2) to perform pose recognition on the cow and obtain the cow skeleton sequence. The steps include:

[0092] 3.1.1) Input 150 videos each of the remaining cows walking normally and limping into the cow posture estimation network;

[0093] 3.1.2) Iterative training using a deep neural network is used to predict the positions of key points on the cow's body in all video frames;

[0094] 3.1.3) Calculate the mean square error between the manually marked joint positions in step 2.3.1) and the network prediction results in step 2.3.2) to measure the effectiveness of the network;

[0095] 3.2) Record the skeletal sequences of the cows walking normally and lamely, respectively;

[0096] 3.3) Extract the hoof data from the four segments of the 300 skeletal sequence, and then use this data to calculate the relative distances between the cow's forelimbs and hindlimbs; let the positions of the cow's left forelimb, right forelimb, left hindlimb, and right hindlimb be x1, x2, y1, and y2, respectively; in this invention, the formula for calculating the relative distance between the hooves is:

[0097]

[0098]

[0099] 3.4) For the relative distances between the forelimbs and hindlimbs obtained in step 2.3.3), construct 300 feature vectors with a dimension of 2*n;

[0100] In step 4):

[0101] 4.1) For the 300 feature vectors in step 3.4), set the labels for normal walking and lameness of the cows to 0 and 1 respectively; 0 represents normal walking of the cows and 1 represents lameness of the cows.

[0102] 4.2) For the 300 feature vectors from step 4.1), divide them into a validation set and a test set in a 7:3 ratio;

[0103] 4.3) The 210 validation set feature vectors obtained in step 4.2) are fed into the Bi-LSTM classifier for training to identify whether the cow is lame and obtain the final classification result.

[0104] Furthermore:

[0105] In step 2.2.1), 2000 side-view images of cows walking were manually marked using Labelme software. A total of 17 cow joints were marked, numbered 1-17, which are the cow's mouth, forehead, neck, spine, tailbone, front left leg root, front left knee, front left hoof, front right leg root, front right knee, front right hoof, rear left leg root, rear left knee, rear left hoof, rear right leg root, rear right knee, and rear right hoof.

[0106] In step 2.2.2), the model weights trained on the COCO dataset are used to learn the posture features of dairy cows through transfer learning, so as to perform target detection and joint detection of dairy cows.

[0107] In step 2.2.3), the cow image is processed by the trained ResNet101 cow pose estimation network to generate a feature image. The most suitable candidate box is extracted from several recommendation boxes using the non-maximum suppression method. For each feature of a fixed size, the network generates a cow target detection box that does not interfere with each other. Based on the detection of the cow target detection box, the trained ResNet101 cow pose estimation network is used to detect the cow's body joints.

[0108] In step 2.2.3), manually labeled cow images are first input into the network. Using a single image as input, the network is iteratively trained through a ResNet101-layer deep convolutional neural network consisting of 3×3 convolutional blocks, 2×2 max pooling layers, and softmax layers to produce a set of confidence maps (two-dimensional probability distributions) describing the position of each key point in the input as output.

[0109] The global maximum value in each confidence graph represents the network's prediction of the location of that joint, thus predicting the location of the cow's body joints in all video frames;

[0110] In step 4.2), the 300 feature vectors are divided into a validation set and a test set in a 7:3 ratio; the validation set contains 105 feature vectors each for normally walking and lame cows, and the test set contains 45 feature vectors each for normally walking and lame cows.

[0111] In step 4.3), 210 feature vectors are fed into the Bi-LSTM classifier for training. The input dimension is 2, and 5 vectors are input into the network each time. The LSTM network contains 150 hidden units, with a maximum of 500 iterations. The output is a classification number of 2. The final classification result is obtained.

[0112] In this embodiment, in order to address the shortcomings of model training difficulty and poor generalization ability caused by insufficient data, the cow pose estimation network utilizes the model weights trained on the COCO dataset through transfer learning.

[0113] To construct a walking motion skeleton map based on dairy cows, a cow pose estimation network was trained using transfer learning to predict joint positions, which greatly reduced the amount of computation and improved the computation speed.

[0114] By integrating a cow pose estimation network and a Bi-LSTM model based on the cow's walking joint skeleton, the determination of whether a cow is lame not only reduces computational complexity but also greatly improves computational speed and accuracy.

[0115] In this embodiment, in step 2), the side view image of the cow is first used as the input image. The classic convolutional network idea of ​​ResNet101 is used to reconstruct the cow pose estimation network. Then, the feature vector of the relative distance between the cow's forelimbs and hindlimbs is obtained by using the identified cow skeleton sequence, so as to identify and classify whether the cow is walking normally or lame.

[0116] Step 4.3) uses Bi-LSTM as a classifier, which can better handle sequence problems.

[0117] The following description, in conjunction with the accompanying drawings, further illustrates this embodiment:

[0118] Currently, most cow posture recognition methods indirectly detect lameness by attaching sensors to cows and obtaining their physical characteristics. This method not only disrupts the natural growth of cows, but also, image-based posture recognition is mostly designed for laboratory animals and poultry, lacking universality and having low accuracy. In contrast, the cow lameness recognition method based on cow posture estimation of this invention is applicable to other large four-limbed animals. It can be retrained by modifying the posture recognition network, making it universal; and it has an accuracy of over 90%, demonstrating feasibility.

[0119] like Figure 1 As shown, this embodiment proposes a method for identifying lameness in dairy cows based on cow posture estimation. It uses a cow posture estimation network and a Bi-LSTM classifier network to identify the cow's walking posture. Specifically, the cow posture recognition method in this example includes the following steps:

[0120] Step 1, as follows Figure 2 As shown, a camera was fixed to the outside of the milking aisle to capture side-view walking videos of the cows. The camera was about 2.5 meters away from the milking aisle, and it was ensured that the cow could walk for at least two movement cycles within one screen of the camera (one movement cycle is defined as the cow raising and lowering all four hooves once). For the cow walking videos captured in the natural living environment, a large number of redundant and similar cow images were removed by using the principle of selecting one image every 10 frames, and representative keyframes that can characterize the cow's posture features were extracted.

[0121] Step 2: Construct a cow pose estimation network based on videos of cows walking in their natural habitat. The specific steps are as follows:

[0122] Step 2.1: In order to reduce the requirements for large-scale datasets and optimize the required model information, the model weights trained on the COCO dataset are used to learn the posture features of dairy cows through transfer learning, so as to more accurately identify the posture of dairy cows.

[0123] Step 2.2: The training image is processed by the trained ResNet101 to generate feature images. The most suitable candidate box is extracted from several recommendation boxes using the non-maximum suppression method. The network generates non-interfering cow target detection boxes for each fixed-size feature.

[0124] The structure diagram of the ResNet101 convolutional neural network is as follows: Figure 7 As shown, the network input is (3,224,224), or (C,H,W), specifically representing the number of channels, height, and width, and the output is (num_classes,7,7). Here, conv stands for convolution, 7×7 refers to the kernel size, 64 refers to the number of kernels (i.e., the number of channels output by this convolutional layer), and / 2 indicates a stride of 2. 3->64 means the number of channels is changed from 3 to 64 through convolution or pooling. maxpool represents a max pooling layer, resulting in an output of (64,56,56). avgpool represents a global average pooling layer, and fc represents a fully connected layer, which changes the number of channels from 2048 to num_classes, where num_classes is the number of classification targets required for this task.

[0125] ResNet101 consists of four layers, each containing two BottleNeck structures: one with the same number of input and output channels, and the other with different numbers of input and output channels.

[0126] The number of input and output channels is the same: Taking the dashed box part of layer 1 in the figure as an example, let the input with shape (C,H,W) be x, and let the three convolutional blocks on the left be functions F(x). The sum of the two (F(x)+x) results in an output with shape (C,H,W), which corresponds to the case where the number of input x and output F(x) channels is the same.

[0127] Different input and output channel numbers: Taking the solid-lined part of layer 1 in the diagram as an example, this structure has an additional convolutional layer on the right, denoted as function G(x). This corresponds to the case where the input x and the output F(x) have different channel numbers. The convolutional layer added on the right transforms x into G(x), and G(x) and F(x) have the same number of channels, which plays a role in matching the difference in input and output dimensions. Therefore, F(x) + G(x) can be summed to obtain an output of the same dimension.

[0128] Step 2.3, a schematic diagram of cow posture recognition is shown below. Figure 3As shown, after generating the cow object detection boxes, the `ploy` function from Python's matplotlib image library is used to generate a pose estimation map of the cow image. The generated image contains the cow object detection boxes, the cow's various joints, and the lines connecting the joints, as shown below. Figure 4 As shown.

[0129] Step 3: The remaining 300 cow videos are sequentially fed into the trained cow pose estimation network. Based on the detected cow target bounding boxes, the trained ResNet101 cow pose estimation network is then used to detect the cow's body joints. The specific steps are as follows:

[0130] 3.1) Record the skeletal sequences of the cows walking normally and lamely, respectively;

[0131] 3.2) Extract the hoof data from the four hooves of the 300 skeletal sequences, and then use these data to calculate the relative distances between the forelimbs and hindlimbs of the dairy cows. Figure 5 This invention presents the curves showing the relative distance changes between the forelimbs and hindlimbs of a dairy cow during normal walking and lameness, wherein... Figure 5 (a) shows the curves illustrating the changes in the relative distance between the forelimbs and hindlimbs of a dairy cow during normal walking. Figure 5 (b) is the curve showing the change in the relative distance between the forelimbs and hindlimbs of a lame cow.

[0132] 3.3) For the relative distances between the forelimbs and hindlimbs obtained in step 3.2), construct 300 2*n feature vectors;

[0133] Step 4, identify the limp classification diagram as shown below. Figure 5 As shown, for videos of cows walking in their natural habitat, the 300 feature vectors in step 3.3) are labeled with 0 for normal walking and 1 for lameness; 0 represents normal walking and 1 represents lameness. The dataset is divided into a validation set and a test set in a 7:3 ratio. The validation set contains 105 feature vectors each for normally walking and lame cows, while the test set contains 45 feature vectors each for both. These are then fed into a Bi-LSTM classifier network for training. The parameters of the Bi-LSTM classifier network are as follows: input dimension is 2, 5 inputs are fed into the network each time, the LSTM network contains 150 hidden units, the maximum number of iterations is 500, and the output classification is 2. The trained Bi-LSTM classifier network is used to identify whether a cow is lame, obtaining the final classification result, as shown below. Figure 6 As shown.

[0134] The above are merely preferred embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for recognizing a cow lameness based on a cow posture estimation, characterized by, The method comprises the following steps: acquiring a walking side view image of a cow; constructing a cow posture estimation model, comprising: constructing a data set, wherein the data set comprises a normal walking side view image of a cow and a lame walking side view image of a cow; labeling the data set; constructing a ResNet101 convolutional neural model; training the ResNet101 convolutional neural model based on the labeled data set to obtain the cow posture estimation model; inputting the walking side view image into the cow posture estimation model to obtain a feature vector, wherein the feature vector comprises a cow forelimb feature vector and a cow hindlimb feature vector; acquiring the feature vector comprises: inputting the walking side view image into the cow posture estimation model to obtain a cow skeleton sequence, wherein the skeleton sequence comprises a normal walking skeleton sequence and a lame walking skeleton sequence; based on the skeleton sequence, acquiring a cow hoof relative distance, wherein the cow hoof relative distance is the relative distance of the forelimb and the hindlimb; based on the cow hoof relative distance, constructing the feature vector; identifying the feature vector to obtain a result of whether the cow has lameness. 2.The cow lameness recognition method based on cow pose estimation according to claim 1, wherein, labeling the data set comprises: labeling the cow body and the joint nodes in the cow walking side view image in the data set, wherein the joint nodes comprise a cow muzzle, a forehead, a neck, a spine, a tail vertebra, a front left leg root, a front left knee, a front left hoof, a front right leg root, a front right knee, a front right hoof, a rear left leg root, a rear left knee, a rear left hoof, a rear right leg root, a rear right knee, and a rear right hoof. 3.The cow lameness recognition method based on cow pose estimation according to claim 1, wherein, acquiring the cow skeleton sequence comprises: inputting the walking side view image into the cow posture estimation model to obtain a feature image; using a non-maximum suppression method, acquiring a cow target detection frame based on the feature image; identifying the cow target detection frame through the cow posture estimation model to obtain each joint node of the cow and the connection line between the joint nodes; based on the joint nodes and the connection line between the joint nodes, acquiring the skeleton sequence. 4.The cow lameness recognition method based on cow pose estimation according to claim 1, wherein, the cow hoof relative distance is: Wherein, x1, x2, y1, y2 are the positions of the left front hoof, the right front hoof, the left rear hoof and the right rear hoof of the cow respectively, , are the relative distances of the front limbs and the rear limbs of the cow respectively. 5.The cow lameness recognition method based on cow pose estimation according to claim 1, wherein, the ResNet101 convolutional neural model comprises a 3x3 size convolution block, a 2x2 size maximum pooling layer, and a softmax layer. 6.The cow lameness recognition method based on cow pose estimation according to claim 1, wherein, training the ResNet101 convolutional neural model comprises: inputting the labeled data set into the ResNet101 convolutional neural model to obtain a confidence map of each joint node position; based on the global maximum representative network in the confidence map, predicting the position of the joint node to obtain the position of the cow body and the joint nodes.

Images