Beef cattle climbing behavior monitoring method and system

Abstract

The invention relates to the technical field of beef cattle breeding monitoring, in particular to a beef cattle climbing behavior monitoring method and system. According to the technical scheme, the method comprises the steps of collecting visual data, sound data and sensor data of beef cattle; processing the collected visual and sound data and sensor data; corresponding features are extracted from the preprocessed visual data, sound data and sensor data, and multi-modal features are obtained; fusing the obtained multi-modal features, and constructing a space-time diagram convolutional network to carry out beef cattle climbing behavior recognition; and constructing a space-time diagram convolutional network: taking space correlation and time sequence information of data into consideration, and capturing complex modes and dynamic changes of beef cattle climbing and crossing behaviors. Through multi-modal data acquisition, effective preprocessing, comprehensive feature extraction and accurate behavior recognition, the beef cattle climbing behavior monitoring accuracy and reliability are improved, the breeding efficiency and the management level are improved, and powerful support is provided for intelligent development of the beef cattle breeding industry.

Description

Technical Field

[0001] The present invention relates to the technical field of beef cattle breeding monitoring, and particularly relates to a method and system for monitoring the mounting behavior of beef cattle. Background Art

[0002] In the beef cattle breeding industry, it is very necessary to accurately monitor the mounting behavior. The mounting behavior is not only closely related to the estrus cycle of beef cattle. Precise grasping of breeding during the estrus period can improve the breeding efficiency and breeding income. Moreover, abnormal changes in the mounting behavior can reflect the health status of beef cattle. Diseases, stress, etc. may all cause changes in behavior. Therefore, effective monitoring of the mounting behavior of beef cattle can improve the breeding success rate of beef cattle and can strengthen the health management level of beef cattle, which has important practical application value.

[0003] The existing beef cattle mounting behavior monitoring technologies have various deficiencies: in data monitoring, some methods rely on single-modal data. Only using visual data is vulnerable to the influence of light and occlusion, and it is difficult to insight into the internal physiological state of beef cattle; only relying on sensor data is prone to misjudgment in the face of complex behavior scenarios; only through sound data is vulnerable to environmental noise interference, which will reduce the recognition accuracy. In algorithm application, traditional machine learning methods require manual feature extraction, which is difficult to comprehensively describe complex dynamic changes. Single deep learning models have the defects of insufficient mining of time series information and prone to gradient problems when dealing with long sequence data.

[0004] Most of the existing monitoring methods do not fully consider the internal correlation and complementarity of multi-modal data, and fail to effectively fuse various data such as vision, sound, and sensors, and cannot comprehensively and accurately describe the mounting behavior of beef cattle; at the same time, traditional methods ignore the spatio-temporal correlation information of data and are difficult to capture the dynamic process of the mounting behavior, resulting in inaccurate monitoring. In view of the many deficiencies of the existing technology and the inability to meet the requirements of accurate and reliable monitoring in actual breeding, a method and system for monitoring the mounting behavior of beef cattle that can improve the accuracy and reliability of monitoring are proposed. Summary of the Invention

[0005] The object of the present invention is to address the problems in the background art and propose a method and system for monitoring the mounting behavior of beef cattle that can improve the accuracy and reliability of monitoring.

[0006] In a first aspect, the present application provides a method for monitoring the mounting behavior of beef cattle, including the following steps:

[0007] Data acquisition: Use a high-definition camera, a high-sensitivity microphone, and an intelligent collar integrated with an accelerometer, a gyroscope, and a pressure sensor to collect visual data, sound data, and sensor data of beef cattle respectively;

[0008] Data preprocessing: Process the collected visual, sound data, and sensor data to improve the data quality;

[0009] Feature extraction: Extract corresponding features from the preprocessed visual data, sound data, and sensor data to obtain multi-modal features;

[0010] Behavior recognition: Fuse the obtained multi-modal features and construct a spatio-temporal graph convolutional network for beef cattle mounting behavior recognition, where:

[0011] Multi-modal feature fusion: Fuse the data of three modalities, namely vision, sound, and sensors, to describe the behavior characteristics of beef cattle from multiple perspectives;

[0012] Construct a spatio-temporal graph convolutional network: Consider the spatial correlation and time series information of the data to capture the complex patterns and dynamic changes of beef cattle mounting behavior.

[0013] Optionally, in the data collection step, an intelligent collar integrating an accelerometer, a gyroscope, and a pressure sensor is used to collect sensor data.

[0014] Optionally, in the data preprocessing step, the visual data is enhanced using the histogram equalization algorithm. The formula is:

[0015]

[0016] where L(x, y) is the gray value of the original image at the coordinate (x, y), L min and L max are the minimum and maximum gray values of the original image respectively, and L’(x, y) is the gray value of the enhanced image at this coordinate; The Gaussian mixture model is used for background subtraction.

[0017] Optionally, in the data preprocessing step, the sound data is filtered to remove environmental noise using a band-pass filter. The frequency range is set at 20 Hz - 20 kHz. After frame processing, each frame is processed with a Hamming window. The formula is:

[0018]

[0019] where n is the sample serial number and N is the length of the frame.

[0020] Optionally, in the data preprocessing step, the sensor data is denoised using a moving average filtering algorithm. The calculation formula is:

[0021]

[0022] where y(t) is the filtered data, x(i) is the original sensor data, M is the length of the sliding window, and t is the current time.

[0023] Optionally, in the feature extraction step, visual features are extracted using a pre-trained ResNet-50 model, and an attention mechanism module is introduced in the last few convolutional layers of the model. The calculation formula for the attention weight α is as follows:

[0024]

[0025] where s is the score of the position, which is calculated through convolutional layers and fully connected layers. In practical applications, the enhanced image is first input into the ResNet-50 model to obtain a feature map, then the feature map is input into the attention mechanism module to calculate the score s of each position, obtaining the attention weight α, and the feature map is weighted according to the attention weight to highlight the key parts and action change areas of the beef cattle.

[0026] Optionally, in the feature extraction step, the Mel-frequency cepstral coefficients are extracted for sound features, and the audio frame is converted to the Mel-frequency scale. The formula is:

[0027]

[0028] where f is the frequency, the energy distribution is calculated on the Mel-frequency scale, and the Mel-frequency cepstral coefficients are obtained through discrete cosine transform.

[0029] Optionally, in the feature extraction step, sensor features are extracted from the data of accelerometers, gyroscopes, and pressure sensors for time-domain features and frequency-domain features. The mean μ, variance σ 2 , peak value x peak of the time-domain features are calculated using the following formulas respectively:

[0030]

[0031] x peak = max(|x(i)|)

[0032] The frequency-domain features are calculated for the power spectral density through fast Fourier transform. The formula is:

[0033]

[0034] where X(f) is the Fourier transform of the signal and N is the signal length.

[0035] Optionally, in the behavior recognition step, the graph convolution operation formula of the spatio-temporal graph convolutional network is:

[0036] y = σ(A · X · W)

[0037] Among them, y is the output feature, σ is the activation function, A is the adjacency matrix representing the correlation relationship between different modality data, constructed by calculating the cosine similarity between different modality features, X is the input feature, W is the weight matrix, and the formula for the long short-term memory network structure is:

[0038] Input gate: i t = σ(W ii x t + W hi h t-1 + b i )

[0039] Forget gate: f t = σ(W if x t + W hf h t-1 + b f )

[0040] Output gate: o t = σ(W io x t + W ho h t-1 + b o )

[0041] Memory cell: c t = f t · c t-1 + i t · tanh(W ic x t + W hc h t-1 + b c )

[0042] Hidden state: h t = o t · tanh(c t )

[0043] Among them, x t is the input at the current moment, h t-1 is the hidden state at the previous moment, c t-1 is the memory cell at the previous moment, W is the weight matrix, b is the bias vector, and σ is the activation function.

[0044] On the other hand, the present application provides a beef cattle mounting behavior monitoring system, including:

[0045] Data acquisition module: acquiring visual data, sound data, and sensor data of beef cattle;

[0046] Data preprocessing module: processing the acquired visual, sound, and sensor data to improve data quality;

[0047] Feature extraction module: Extract corresponding features from the preprocessed visual data, sound data and sensor data to obtain multi-modal features;

[0048] Behavior recognition module: Fuse the obtained multi-modal features, construct a spatio-temporal graph convolutional network for beef cattle mounting behavior recognition, where:

[0049] Multi-modal feature fusion: Fuse the data of three modalities, namely vision, sound and sensors, to describe the behavior characteristics of beef cattle from multiple angles;

[0050] Construct a spatio-temporal graph convolutional network: Consider the spatial correlation and time series information of the data to capture the complex patterns and dynamic changes of beef cattle mounting behavior.

[0051] Compared with the prior art, the present application includes at least one of the following beneficial technical effects:

[0052] The present invention adopts a multi-modal all-round acquisition technology to achieve no monitoring blind area, can sensitively capture the subtle behavior changes of beef cattle, the device can also adapt to various breeding environments, ensure the stability and reliability of data, and the behavior recognition uses a spatio-temporal graph convolutional network to effectively capture the complex behavior patterns and dynamic changes of beef cattle; further, through accurate data acquisition and behavior recognition, scientific intervention is carried out on the breeding and reproduction process of beef cattle, the reproduction efficiency is improved, the health status is monitored in real time, the health management level can be improved, and then the breeding risk is reduced, providing strong support for the intelligent development of the beef cattle breeding industry. Brief description of the drawings

[0053] Figure 1 It is a flowchart of a method for monitoring beef cattle mounting behavior;

[0054] Figure 2 It is a hardware block diagram of a system for monitoring beef cattle mounting behavior. Detailed implementation manners

[0055] The technical solutions of the present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0056] Embodiment

[0057] As Figure 1 - Figure 2 shown, a method for monitoring beef cattle mounting behavior proposed by the present invention includes the following steps:

[0058] Data acquisition: Use a high-definition camera, a high-sensitivity microphone, and an intelligent collar integrated with an accelerometer, a gyroscope, and a pressure sensor to collect visual data, sound data, and sensor data of beef cattle respectively;

[0059] Data preprocessing: Process the collected visual, sound data and sensor data to improve the data quality;

[0060] Feature extraction: Extract corresponding features from the preprocessed visual data, sound data, and sensor data to obtain multi-modal features;

[0061] Behavior recognition: Fuse the obtained multi-modal features and construct a spatio-temporal graph convolutional network for beef cattle mounting behavior recognition, where:

[0062] Multi-modal feature fusion: Fuse the modal data of vision, sound, and sensors to describe the behavior characteristics of beef cattle from multiple angles;

[0063] Construct a spatio-temporal graph convolutional network: Capture the complex patterns and dynamic changes of beef cattle mounting behavior according to the spatial correlation and time series information of the data.

[0064] Furthermore, in the data collection step, high-definition cameras are installed at the four corners and the central position at the top of the cattle shed, installed at a 45-degree downward angle, with a frame rate set to 30fps and a resolution of 1920×1080 to ensure that the subtle movements of beef cattle can be clearly captured; high-sensitivity microphones are evenly distributed on the ceiling of the cattle shed, with a sampling frequency of 44.1kHz and a quantization bit number of 16 bits, used to collect the sounds generated by beef cattle activities in all directions; intelligent collars are worn around the necks of each beef cattle, and the accelerometer and gyroscope collect the acceleration and angular velocity data of the beef cattle in the X, Y, and Z axes at a frequency of 100Hz, and the pressure sensor monitors the pressure change of the contact between the beef cattle's neck and the collar in real time, indirectly reflecting the body movements and posture changes of the beef cattle.

[0065] Secondly, in the data preprocessing step, the visual data is enhanced using the histogram equalization algorithm, and the formula is:

[0066]

[0067] The above formula is used to remap the gray values of the image. L(x,y) is the gray value of the original image at the coordinate (x,y), and L min and L max are the minimum and maximum gray values of the original image respectively, and L’(x,y) is the gray value of the enhanced image at this coordinate; the Gaussian mixture model is used for background subtraction. Through remapping, the contrast and clarity of the image can be improved, making the bright parts of the image brighter and the dark parts darker, highlighting the details of the image;

[0068] In the specific usage stage, for each collected visual image, traverse each pixel point in the image to obtain its gray value L(x,y), and at the same time find the minimum gray value L min and the maximum gray value L max, and then substitute it into the above formula to calculate the enhanced grayscale value L’(x,y), and assign the calculated grayscale value to the corresponding pixel points, thus completing the enhancement processing of the entire image. Then, the Gaussian mixture model is used for background subtraction to separate the dynamic contour of the beef cattle from the static background.

[0069] Secondly, in the data preprocessing step, the environmental noise is removed from the sound data by using a band-pass filter, and the frequency range is set at 20Hz - 20kHz, only retaining the sound frequency band related to the behavior of beef cattle; then, frame division is performed with 25ms as one frame, each frame overlapping by 10ms, and a Hamming window is added to each frame. The formula is:

[0070]

[0071] The above formula is used to generate the Hamming window function. n is the sample serial number, N is the length of the frame. Adding the Hamming window processing can reduce spectral leakage, make the boundary of the audio frame smoother, and more accurately reflect the frequency characteristics of the sound;

[0072] Specifically, when in use, for each audio frame after frame division, according to the length N of the frame, calculate the window function value w(n) corresponding to each sample serial number n according to the above formula, and multiply each sample value in the audio frame by the corresponding window function value to obtain the windowed audio frame. For example: if the length N of the audio frame = 1000, for the sample serial number n = 10, first calculate Then calculate the cosine value, and finally substitute it into the formula to get w(10), and multiply the 10th sample value in the audio frame by w(10).

[0073] Secondly, in the data preprocessing step, the sensor data is denoised by using a moving average filtering algorithm. For a moving window with a length of M, the calculation formula for the filtered data y(t) at the current moment t is:

[0074]

[0075] The above formula eliminates the high-frequency noise interference in the data by averaging the original sensor data within the moving window. x(i) is the original sensor data, M is the length of the moving window, t is the current moment, and y(t) is the filtered data;

[0076] In the specific usage stage, first determine the length M of the sliding window. For the data sequence collected by the sensor, starting from the M-th data point, calculate the average value of the data at the current time t and the previous M - 1 times as the filtered data y(t); for example, if the length M of the sliding window is 5 and the current time t is 10, then calculate the average value of the 5 data from x(6) to x(10) as y(10); as time goes by, the sliding window moves backward continuously, and calculate the filtered data at each time in turn to make the data smoother and more stable.

[0077] Furthermore, in the feature extraction step, the visual features are extracted using a pre-trained ResNet-50 model, and an attention mechanism module is introduced in the last few convolutional layers of the model. The calculation of the attention weight α is as follows:

[0078]

[0079] The above formula is used to calculate the attention weight. s is the score of the position, which is calculated through convolutional layers and fully connected layers, ∑ i exp(s i ) is the sum of the exponential values of all position scores. The attention weight α represents the importance of each position in feature extraction. The higher the score of a position, the greater its corresponding attention weight;

[0080] In specific use, first input the visually enhanced image into the pre-trained ResNet-50 model, obtain the feature map through a series of convolutional operations, then input the feature map into the attention mechanism module, calculate the score s of each position through convolutional layers and fully connected layers. For each position, calculate the exponential value exp(s) of its score, and at the same time calculate the sum of the exponential values of all position scores ∑ i exp(s i )), and finally divide the exp(s) of this position by ∑ i exp(s i ) to obtain the attention weight α; weight the feature map according to the calculated attention weight to highlight the key parts and action change areas of the beef cattle and improve the accuracy of feature extraction.

[0081] Secondly, in the feature extraction step, the Mel Frequency Cepstral Coefficients (MFCCs) are extracted for the sound features. First, convert the audio frames to the Mel frequency scale. The formula is:

[0082]

[0083] The above formula converts the linear frequency f to the Mel frequency mel(f), which is more in line with the perceptual characteristics of the human auditory system for sound frequencies. Since the human auditory system has different sensitivities to different frequencies, the Mel frequency scale can better reflect this difference. f is the frequency, and the energy distribution is calculated on the Mel frequency scale, and the Mel frequency cepstral coefficients are obtained through the discrete cosine transform;

[0084] In the specific usage process, for each audio frame after frame division and windowing, first perform a fast Fourier transform to obtain its spectrum, obtain each frequency component f in the spectrum, substitute each frequency f into the above formula, and calculate the corresponding Mel frequency mel(f); for example, if a certain frequency f = 1000 Hz, then calculate take its common logarithm and multiply by 2595 to obtain the corresponding Mel frequency, then calculate the energy distribution on the Mel frequency scale, and obtain the Mel frequency cepstral coefficients through the discrete cosine transform for subsequent sound feature analysis.

[0085] Secondly, in the feature extraction step, the sensor features extract time-domain features and frequency-domain features from the accelerometer, gyroscope, and pressure sensor data. The mean μ and variance σ of the time-domain features 2 , peak value x peak The calculation formulas are as follows:

[0086]

[0087] x peak = max(|x(i)|)

[0088] The frequency-domain features calculate the power spectral density through the fast Fourier transform, and the formula is:

[0089]

[0090] where:

[0091] The formula for the mean μ: used to calculate the average level of the sensor data sequence, reflecting the central tendency of the data. x(i) is the i-th data in the sensor data sequence, and N is the length of the data sequence;

[0092] The variance σ 2 Formula: used to measure the degree of dispersion of the sensor data sequence, that is, the dispersion of the data relative to the mean;

[0093] The formula for the peak value x peak Formula: used to find the maximum absolute value in the sensor data sequence, reflecting the maximum fluctuation amplitude of the data;

[0094] Power spectral density \(P(f)\) formula: The frequency-domain representation \(x(f)\) is obtained by performing a fast Fourier transform on the sensor data. The square of its modulus is calculated and divided by the data length \(N\) to obtain the energy distribution of the signal at different frequencies;

[0095] The specific usage method of the above formula is as follows:

[0096] Mean calculation: For a sensor data sequence of length \(N\), all the data in the sequence are added together and then divided by \(N\) to obtain the mean \(\mu\);

[0097] Variance calculation: First, calculate the mean \(\mu\). Then, for each data \(x(i)\) in the sequence, calculate the square of the difference between it and the mean \((x(i)-\mu)\) 2 , and add the squared values and divide by \(N\) to obtain the variance \(\sigma\) 2 ;

[0098] Peak calculation: Traverse the sensor data sequence, find the absolute value of each data, and take the maximum value as the peak \(x\) peak ;

[0099] Power spectral density calculation: Perform a fast Fourier transform on the sensor data sequence to obtain \(x(f)\), calculate the square of its modulus \(|x(f)|\) 2 , and then divide by the data length \(N\) to obtain the power spectral density \(P(f)\), which comprehensively describes the motion state and body force conditions of beef cattle through these features.

[0100] It should be noted that in the behavior recognition step, the graph convolution operation formula of the spatio-temporal graph convolutional network is:

[0101] \(y = \sigma(A\cdot X\cdot W)\)

[0102] In the above formula, \(y\) is the output feature, \(\sigma\) is the activation function, \(A\) is the adjacency matrix, which represents the correlation relationship between different modal data and is constructed by calculating the cosine similarity between different modal features. \(X\) is the input feature, and \(W\) is the weight matrix. The formula for the long short-term memory network structure is:

[0103] Input gate: \(i\) t \(=\sigma(W\) ii \(x\) t +W\) hi \(h\) t-1 +b\) i )

[0104] Forget gate: \(f\) t \(=\sigma(W\) if \(x\) t +W\) hf \(h\) t-1 +b\) f )

[0105] Output gate: \(o\)t =σ(W io x t +W ho h t-1 +b o )

[0106] Memory unit: c t =f t ·c t-1 +i t tanh(W ic x t +W hc h t-1 +b c )

[0107] Hidden state: h t =o t ·tanh(c t )

[0108] Among them, x t is the input at the current moment, h t-1 is the hidden state of the previous moment, c t-1 is the memory unit of the previous moment, W is the weight matrix, b is the bias vector, and σ is the sigmoid activation function;

[0109] Here is an explanation of the above formula and a list of specific usage methods:

[0110] Graph convolution operation formula: used to mine the potential connections between different modal data. The adjacency matrix A represents the correlation between nodes of different modal data. It is constructed by calculating the cosine similarity between different modal features, reflecting the connection strength between nodes. The input feature X is the fused multimodal feature, and the weight matrix W is the parameter that needs to be learned through training. First, perform matrix multiplication on A, X, and W, and then introduce nonlinear factors through the activation function σ to obtain the output feature y. When used specifically, perform matrix multiplication on the fused multimodal feature X with the pre-built adjacency matrix A and the trained weight matrix W, and then input the result into the activation function (such as: ReLU function, if the result is less than 0, output 0, and if it is greater than 0, output the result itself) to obtain the output feature y.

[0111] Long short-term memory network structure formula:

[0112] Input Gate i t :Based on the current input x t and the hidden state h at the previous moment t-1 , determines how much new information can enter the memory unit at the current moment; through the linear combination W ii x t +W hi ht-1 +b i And it passes through the sigmoid activation function to map the input value between 0 and 1, obtaining the value i of the input gate t ; When specifically used, the input x at the current moment t and the hidden state h at the previous moment t-1 are respectively multiplied by the corresponding weight matrices W ii and W hi , added with the bias vector b i , and then calculated through the sigmoid function to obtain i t ;

[0113] Forget gate f t : Based on the input x at the current moment t and the hidden state h at the previous moment t-1 , it decides how much information in the memory cell c at the previous moment t-1 needs to be forgotten. The calculation method is the same as that of the input gate. Through W if x t +W hf h t-1 +b f and passing through the sigmoid activation function to obtain f t ;

[0114] Output gate o t : Based on the input x at the current moment t and the hidden state h at the previous moment t-1 , it decides how much information in the memory cell c at the current moment t can be output to the hidden state h t , and o t is also calculated through linear combination and the sigmoid activation function;

[0115] Memory cell c t : It decides how much information in the memory cell c at the previous moment t should be retained according to the forget gate f t-1 , and decides how much new information should be added according to the input gate i t . The amount of added information is obtained through:

[0116] tanh(W ic x t +W h xh t-1 +b c ). Adding the two together gives the memory cell c at the current moment t ;

[0117] Hidden state h t : According to the output gate o tDetermine the memory cell c from the current moment t How much information is output to the hidden state h t , through o t Multiply by tanh(c t ) is obtained; in practical applications, the fused feature sequence is input into the long short-term memory network structure in chronological order. As the time step advances, the long short-term memory network structure continuously learns the time variation law of the behavior, thereby accurately identifying the climbing behavior of beef cattle.

[0118] In this embodiment, during the data collection stage, a multi-device and multi-angle data collection method is adopted. High-definition cameras are installed in multiple key locations in the cowshed, and high-sensitivity microphones and smart collars are used to achieve all-round collection of beef cattle's visual, sound and sensor data. This not only ensures that there are no blind spots in monitoring, but also accurately captures the subtle movements, sounds, body posture and movement status changes of beef cattle, providing a rich and accurate data basis for subsequent analysis, achieving the effect of comprehensive coverage and precise capture; the high-definition camera has a starlight-level low-illumination function, which can achieve clear imaging even in dimly lit environments, and the high-sensitivity microphone is optimized through the layout of acoustic simulation software, which effectively reduces the interference of environmental factors on data collection, ensuring that high-quality data can be stably obtained in various breeding environments and can adapt to different environments.

[0119] In the data preprocessing stage, corresponding preprocessing algorithms are used for data of different modalities, such as histogram equalization and background subtraction of visual data, bandpass filtering and windowing processing of sound data, sliding average filtering of sensor data, etc., to effectively remove noise interference, improve the contrast, clarity and stability of the data, and provide more reliable data for subsequent feature extraction; through preprocessing, data of different modalities are made more in line with the requirements of subsequent analysis, the availability and interpretability of the data are improved, which helps to mine the behavioral information hidden in the data and effectively enhances the availability of the data.

[0120] In the feature extraction stage, rich and diverse features are extracted from visual, sound and sensor data, such as appearance, motion trajectory and posture information in visual features, Mel-frequency cepstral coefficients in sound features, time domain and frequency domain features in sensor features, etc. The multi-dimensional features can comprehensively and deeply describe the behavioral characteristics of beef cattle and provide strong support for the accurate identification of mounting behavior. Introducing the attention mechanism in visual feature extraction can enable the model to pay more attention to the key parts and movement changes of beef cattle, highlight the important information related to mounting behavior, and improve the accuracy and pertinence of feature extraction.

[0121] In the behavior recognition stage, behavior recognition is carried out by constructing a spatio-temporal graph convolutional network, which integrates the spatial correlation and time series information of multi-modal data. The graph convolution operation can explore the potential connections between different modal data, and the long short-term memory network structure can learn the temporal variation law of behaviors, so as to more effectively capture the complex patterns and dynamic changes of beef cattle mounting behaviors, with higher recognition accuracy and stronger adaptability compared with traditional methods; by setting a reasonable probability threshold for behavior determination, the situations of misjudgment and missed judgment can be effectively reduced, providing more accurate decision-making basis for breeders, timely discovering the mounting behaviors of beef cattle, grasping the best breeding opportunity, and improving the breeding success rate.

[0122] To sum up, accurately and timely monitoring the mounting behaviors of beef cattle helps breeders reasonably arrange breeding plans, improve breeding efficiency, increase breeding income, and can also timely discover the abnormal behaviors of beef cattle, provide references for health management, and reduce breeding risks; the innovation and effectiveness of this method provide new ideas and technical support for the intelligent development of the beef cattle breeding industry, contribute to improving the breeding management level of the whole industry, and promoting the development of the beef cattle breeding industry towards precision and intelligence; through multi-modal data collection, effective preprocessing, comprehensive feature extraction and accurate behavior recognition, the present invention not only improves the accuracy and reliability of beef cattle mounting behavior monitoring, but also enhances breeding efficiency and management level, providing strong support for the intelligent development of the beef cattle breeding industry.

[0123] The above specific embodiments are only several alternative embodiments of the present invention. Based on the technical solution of the present invention and the relevant inspirations of the above embodiments, those skilled in the art can make various alternative improvements and combinations to the above specific embodiments.

Claims

1. A monitoring method for the mounting behavior of beef cattle, characterized in that, It includes the following steps: Data acquisition: Collect visual data, sound data, and sensor data of beef cattle; Data preprocessing: Process the collected visual, sound data, and sensor data; Feature extraction: Extract corresponding features from the preprocessed visual data, sound data, and sensor data to obtain multi-modal features; Behavior recognition: Fuse the obtained multi-modal features, construct a spatio-temporal graph convolutional network for beef cattle mounting behavior recognition, where: Multi-modal feature fusion: Fuse the modal data of vision, sound, and sensors to describe the behavior characteristics of beef cattle from multiple angles; Construct a spatio-temporal graph convolutional network: Capture the complex patterns and dynamic changes of beef cattle mounting behavior according to the spatial correlation and time series information of the data.

2. The beef cattle mounting behavior monitoring method according to claim 1, wherein In the data acquisition step, an intelligent collar integrating an accelerometer, a gyroscope, and a pressure sensor is used to collect sensor data.

3. A beef cattle mounting behavior monitoring method according to claim 1, characterized in that, In the data preprocessing step, the visual data is enhanced by the histogram equalization algorithm, and the formula is: Among them, L(x,y) is the gray value of the original image at the coordinate (x,y), L min and L max are the minimum and maximum grayscale values of the original image, respectively, and L'(x, y) is the grayscale value of the enhanced image at this coordinate; the Gaussian mixture model is used for background subtraction.

4. A method for monitoring the mounting behavior of beef cattle according to claim 1, characterized in that In the data preprocessing step, the sound data is processed by a band-pass filter to remove environmental noise. After frame division processing, each frame is processed with a Hamming window, and the formula is: where n is the sample serial number and N is the length of the frame.

5. A beef cattle mounting behavior monitoring method according to claim 1, characterized in that, In the data preprocessing step, the sensor data is denoised by a moving average filtering algorithm, and the calculation formula is: where y(t) is the filtered data, x(i) is the original sensor data, M is the length of the moving window, and t is the current time.

6. The beef cattle mounting behavior monitoring method according to claim 1, characterized in that In the feature extraction step, the visual features are extracted by a pre-trained ResNet-50 model, and an attention mechanism module is introduced into the last few convolutional layers of the model. The calculation formula of the attention weight α is: where s is the score of the position, which is calculated by the convolutional layer and the fully connected layer. First, the enhanced image is input into the ResNet-50 model to obtain the feature map, and then the feature map is input into the attention mechanism module to calculate the score s of each position to obtain the attention weight α. The feature map is weighted according to the attention weight to highlight the key parts and action change areas of the beef cattle.

7. A method for monitoring the mounting behavior of beef cattle according to claim 1, characterized in that In the feature extraction step, the Mel frequency cepstral coefficients are extracted for the sound features, and the audio frame is converted to the Mel frequency scale. The formula is: where f is the frequency, and the energy distribution is calculated on the Mel frequency scale. The Mel frequency cepstral coefficients are obtained through discrete cosine transform.

8. A method for monitoring the mounting behavior of beef cattle according to claim 1, characterized in that, In the feature extraction step, time-domain features and frequency-domain features are extracted from the data of the accelerometer, gyroscope, and pressure sensor as the sensor features. The mean value μ and variance σ of the time-domain features 2 , peak value x peak are calculated by the following formulas respectively: x peak = max(|x(i)|) The frequency domain features are calculated by the fast Fourier transform to obtain the power spectral density, and the formula is: where X(f) is the Fourier transform of the signal and N is the signal length.

9. A method for monitoring the mounting behavior of beef cattle according to claim 1, characterized in that In the behavior recognition step, the graph convolution operation formula of the spatio-temporal graph convolutional network is: y = σ(A·X·W) where y is the output feature, σ is the activation function, A is the adjacency matrix, which represents the correlation relationship between different modal data and is constructed by calculating the cosine similarity between different modal features, X is the input feature, and W is the weight matrix. The formula of the long short-term memory network structure is: Input gate: i t = σ(W ii x t + W hi h t-1 + b i ) Forgotten Gate: f t = σ(W if x t + W hf h t-1 + b f ) Output gate: o t = σ(W io x t + W ho h t-1 + b o ) Memory unit: c t = f t ·c t-1 + i t ·tanh(W ic x t + W hc h t-1 + b c ) Hidden state: h t = o t ·tanh(c t ) where x t is the input at the current time, h t-1 is the hidden state at the previous time, c t-1 is the memory cell at the previous time, W is the weight matrix, b is the bias vector, and σ is the sigmoid activation function.

10. The beef cattle mounting behavior monitoring system according to claim 1, wherein It includes: Data acquisition module: Collect visual data, sound data, and sensor data of beef cattle; Data preprocessing module: Process the collected visual, sound data, and sensor data; Feature extraction module: Extract corresponding features from the preprocessed visual data, sound data, and sensor data to obtain multi-modal features; Behavior recognition module: Fuse the obtained multi-modal features and construct a spatio-temporal graph convolutional network for beef cattle mounting behavior recognition, where: Multi-modal feature fusion: Fuse the modal data of vision, sound, and sensors to describe the behavior characteristics of beef cattle from multiple perspectives; Construct a spatio-temporal graph convolutional network: Capture the complex patterns and dynamic changes of beef cattle mounting behavior according to the spatial correlation and time series information of the data.

Images