Cow oestrus monitoring method based on multi-camera data fusion

Abstract

The invention discloses a cattle oestrus monitoring method based on multi-camera data fusion, and relates to the technical field of cattle oestrus monitoring. According to the cattle oestrus monitoring method based on multi-camera data fusion, cattle behaviors under complete shielding are visualized through multi-camera dynamic cooperation and skeleton movement deduction; a space-time slice attention mechanism is combined with joint dynamics calculation, target actions are reversely deduced from adjacent cattle body movements, the problem of climbing misjudgment caused by shielding of a traditional visual system is thoroughly solved, and the monitoring reliability in dense cattle herds is remarkably improved; the method comprises the following steps: establishing a differential filtering model of limp gaits and oestrus behaviors; based on acceleration signal time-frequency characteristics and duration determination, a false alarm signal is dynamically suppressed, the inherent defects of a contact sensor in a special physiological state are overcome, and the manual re-checking cost is reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of cattle estrus monitoring, and in particular to a cattle estrus monitoring method based on multi-camera data fusion. Background Art

[0002] As dairy cattle farming develops towards intensive production, large-scale ranches with more than 1,000 heads have become the mainstream. Such ranches generally adopt a high-density feeding model, and when cattle gather in feeding areas and rest areas, the distance between individuals is often less than 0.5 meters. The mounting behavior of cows in estrus often occurs in these areas. The current mainstream solution relies on multi-camera vision systems or wearable sensors, replacing manual observation with automated technology to improve the efficiency and timeliness of estrus detection.

[0003] Common monitoring solutions include the introduction of three-dimensional posture estimation and multi-view fusion algorithms, such as cow key point detection based on YOLOv8. However, when an estrus cow is surrounded by multiple cows and completely blocked, the visual system loses the target posture information, and the algorithm misjudges the climbing action as standing or squeezing. In addition, the existing system uses a rotation view switching strategy, and it takes more than 2 seconds from detecting occlusion to calling the backup view, missing the key frame of the climbing behavior. The accelerometers worn on the neck used in some solutions are interfered with by collisions with neighboring cows in dense groups, generating climbing-like signal noise.

[0004] To address this issue, some solutions increase camera density, adding a bird's-eye view camera for every 50 square meters of cattle shed. However, this creates blind spots due to occlusion caused by ceiling brackets. Alternatively, occlusion relationships are inferred based on the distance to the cattle. However, positioning errors greater than 15 cm can amplify misjudgments in densely populated scenarios. These solutions only partially reduce the missed detection rate and have yet to resolve the core issue of behavioral deduction under complete occlusion. Summary of the Invention

[0005] In view of the above existing problems, the present invention is proposed.

[0006] The present invention provides a cattle estrus monitoring method based on multi-camera data fusion to solve the problem of high missed detection and misjudgment rate of estrus behavior caused by occlusion, interference and delay in existing monitoring technology in high-density breeding scenarios.

[0007] In order to solve the above technical problems, the present invention provides the following technical solutions:

[0008] The embodiment of the present invention provides a cattle estrus monitoring method based on multi-camera data fusion, which includes:

[0009] Step S1: deploy at least four wide-angle camera groups on the ceiling of the cowshed, each group including a visible light lens and a thermal imaging lens, and each lens is connected to the edge computing node through a servo pan / tilt;

[0010] Step S2: The edge computing node acquires multiple video streams in real time and constructs a three-dimensional skeletal model database of cattle based on the cattle's back pattern and body contour features;

[0011] Step S3: When the visual occlusion area of ​​the target cow exceeds a preset threshold, the pan / tilt of the adjacent camera group is dispatched to perform visual angle compensation and extract motion vector data of the adjacent cow bodies in the occlusion area;

[0012] Step S4, deducing the joint displacement dynamics of the occluded cow based on the motion vector data to generate a virtual skeleton action sequence;

[0013] Step S5: If the virtual skeleton motion sequence detects the associated motion of the hind limbs pushing off the ground and the neck extending forward, it is marked as a suspected climbing event;

[0014] Step S6: fusing visible light texture and thermal imaging temperature field change data to perform multispectral verification on the suspected climbing event;

[0015] Step S7: outputting the estrus behavior determination result to the pasture management system.

[0016] As a preferred embodiment of the cattle estrus monitoring method based on multi-camera data fusion described in the present invention, step S3 includes:

[0017] The PTZ is scheduled to perform horizontal rotation compensation and pitch angle adjustment. The horizontal rotation range is ±30° and the pitch angle adjustment is 10°-15°.

[0018] At the same time, the thermal imaging lens is activated to assist in positioning, and the temperature field change data is used to assist in spatial positioning.

[0019] As a preferred embodiment of the cattle estrus monitoring method based on multi-camera data fusion described in the present invention, step S4 includes:

[0020] The spatiotemporal slicing attention mechanism is used to analyze the motion trajectories of adjacent cattle, and the skeletal displacement dynamics of the obscured cattle are deduced based on the spatial topological relationship of the herd.

[0021] The joint displacement dynamics include the change in angular velocity of the hind limb joints and the direction of the trunk center of gravity offset.

[0022] As a preferred solution of the cattle estrus monitoring method based on multi-camera data fusion described in the present invention, the dynamic deduction of joint displacement specifically includes:

[0023] a) Extract the displacement acceleration of adjacent bovine knee joints;

[0024] b) Calculate the force transfer coefficient of the blocked cattle based on the average distance between the cattle;

[0025] c) Generate the envelope of the hind limb joint motion trajectory by combining spatial topological constraints.

[0026] As a preferred embodiment of the cattle estrus monitoring method based on multi-camera data fusion described in the present invention, in the step of calculating the change in angular velocity of the hind limb joints in step S4, the coordinates of the hind limb hip-knee-ankle points are extracted from the three-dimensional skeletal model, and the following calculations are performed in sequence:

[0027] Calculate the joint angle:

[0028]

[0029] Among them, θ h (t) represents the angle of the hind limb joint at time t, v thigh (t) represents the vector from the hip joint to the knee joint, represented by p hip (t)-p knee (t) calculation, v shank (t) represents the vector from the knee joint to the ankle joint, represented by p ankle (t)-p knee (t) calculation, p hip (t) represents the spatial coordinate of the cow's hip joint at time t, p knee (t) represents the spatial coordinate of the cattle knee joint at time t, p ankle (t) represents the spatial coordinates of the ankle joint of the cow at time t, |·| represents the Euclidean norm of the vector;

[0030] Calculate the instantaneous angular velocity,

[0031]

[0032] Among them, ω h (t) represents the instantaneous angular velocity of the hind limb joint at time t, θ h (t-Δt) represents the joint angle at the previous frame time t-Δt, θ h In radians, ω h (t) is in radians per second, and Δt represents the time interval between two adjacent frame samples, in seconds;

[0033] Calculate the change in angular velocity:

[0034] Δω h (t) = ω h (t)-ω h (t-Δt),

[0035] Among them, Δω h (t) represents the change in angular velocity of the hind limb joints at time t, ω h (t-Δt) represents the instantaneous angular velocity of the previous frame;

[0036] Exponential smoothing filter:

[0037]

[0038] in, It represents the change in angular velocity after smoothing, α is the exponential smoothing coefficient, dimensionless, ranging from 0.2 to 0.3, and is used to balance response speed and noise suppression. Indicates the smoothed angular velocity change of the previous frame;

[0039] Smoothing the results It is used as the dynamic feature of the hind limbs and is input into the virtual skeleton action sequence generation module.

[0040] As a preferred embodiment of the cattle estrus monitoring method based on multi-camera data fusion described in the present invention, step S5 includes:

[0041] When the acceleration sensor returns the climbing characteristic signal, the fundamental frequency of the gait of the cattle in the previous period is synchronously analyzed;

[0042] If limping walking characteristics are present, a dynamic suppression coefficient is applied to the suspected climbing event, and the suppression coefficient decreases as the duration of the limping characteristics increases.

[0043] As a preferred embodiment of the cattle estrus monitoring method based on multi-camera data fusion described in the present invention, step S6 includes:

[0044] Dynamically allocate the fusion weight of visible light and thermal imaging according to the ambient light intensity;

[0045] Under strong backlight conditions, the thermal imaging temperature gradient map is used to guide the edge restoration of visible light images;

[0046] The edge restoration adopts a generative adversarial network, which takes the temperature field of thermal imaging as a conditional input to generate dust removal texture features;

[0047] The repaired image is input into the improved YOLOv8 model, which incorporates a channel attention mechanism in the feature extraction layer.

[0048] As a preferred embodiment of the cattle estrus monitoring method based on multi-camera data fusion described in the present invention, in step S6, after measuring the ambient light intensity at time t, a fusion weight calculation is performed, including:

[0049] Measure and normalize the ambient light, let:

[0050] I a (t) = S lux (t),

[0051] Among them, Ia (t) represents the ambient light intensity at time t, unit: lux, S lux (t) represents the raw measurement value from the camera’s built-in light sensor;

[0052] Visible light weight map:

[0053]

[0054] in, Represents the uncropped visible light fusion weight, β represents the light response steepness coefficient, unit: lux -1 , I m Indicates the neutral light threshold, unit: lux,

[0055] In the formula, the steepness coefficient β and the neutral threshold I m Determined by the preset critical light range:

[0056]

[0057] Among them, W v,min Represents the minimum visible light weight, W v,max Indicates the highest visible light weight, L low Indicates the low light threshold, unit: lux, L high Indicates the high light threshold, unit: lux,

[0058]

[0059] Clip visible light weights:

[0060]

[0061] Among them, W v (t) represents the visible light fusion weight after clipping, max(·) and min(·) are the element-wise maximum and minimum operations, respectively;

[0062] Calculate the thermal imaging weight:

[0063] W t (t)=1-W v (t),

[0064] Among them, W t (t) represents the thermal imaging fusion weight,

[0065] Pixel-level image fusion:

[0066] I f (x,y,t)=W v (t)I vis (x,y,t)+W t (t)Ith (x,y,t),

[0067] Among them, I f (x, y, t) represents the pixel intensity of the fused image, I vis (x, y, t) represents the visible light image pixel intensity, I th (x, y, t) represents the normalized pixel value of the thermal imaging image, and (x, y) represents the pixel coordinates.

[0068] As a preferred solution of the cattle estrus monitoring method based on multi-camera data fusion described in the present invention, the edge computing node adopts a regional processing architecture, the camera's built-in FPGA performs target detection, and the edge server runs a cross-lens spatiotemporal alignment algorithm;

[0069] The spatiotemporal alignment algorithm uses the three-dimensional skeleton model database as a benchmark to map multiple video streams into a unified coordinate system.

[0070] As a preferred solution of the cattle estrus monitoring method based on multi-camera data fusion described in the present invention, the cross-lens spatiotemporal alignment algorithm includes a skeletal key point matching module, which realizes multi-view data association through the spatial position of the cattle's scapula and hip joint.

[0071] The beneficial effects of the present invention are as follows: the present invention visualizes the behavior of cattle under complete occlusion through dynamic coordination of multiple cameras and skeletal motion deduction; the spatiotemporal slicing attention mechanism is combined with joint dynamics calculation to infer the target action from the motion of adjacent cattle bodies, completely solving the climbing misjudgment problem caused by occlusion in traditional visual systems, and significantly improving the monitoring reliability in dense cattle herds;

[0072] The present invention innovatively establishes a differential filtering model for limping gait and estrus behavior; based on the time-frequency characteristics and duration judgment of acceleration signals, it dynamically suppresses false alarm signals, overcomes the inherent defects of contact sensors under special physiological conditions, and reduces the cost of manual review; introduces illumination weighting function and thermal imaging-guided edge repair technology to constitute a double insurance for enhanced environmental robustness; through logistic mapping and normalization fusion, it automatically enhances the effective information source under strong backlight or dust conditions, ensuring that multispectral data is always in the optimal analysis state, breaking through the failure bottleneck of existing systems in extreme environments.

[0073] The regional processing architecture of the present invention disperses the computing load, and the spatiotemporal alignment algorithm reduces data transmission volume by matching skeletal key points; the three-dimensional model-driven processing flow avoids redundant calculations, allowing complex analysis to be completed locally in real time on the ranch, meeting the needs of large-scale applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0074] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0075] Figure 1 This is a flow chart of the cattle estrus monitoring method based on multi-camera data fusion in Example 1. DETAILED DESCRIPTION

[0076] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0077] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0078] Secondly, the term "one embodiment" or "embodiment" herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in various places throughout this specification does not necessarily refer to the same embodiment, nor does it refer to a separate or selective embodiment that is mutually exclusive of other embodiments.

[0079] Example 1, with reference to Figure 1 This embodiment provides a cattle estrus monitoring method based on multi-camera data fusion, comprising the following steps:

[0080] Step S1: deploy at least four wide-angle camera groups on the ceiling of the cowshed, each group including a visible light lens and a thermal imaging lens, and each lens is connected to the edge computing node through a servo pan / tilt;

[0081] Step S2: The edge computing node acquires multiple video streams in real time and constructs a three-dimensional skeletal model database of cattle based on the cattle's back pattern and body contour features;

[0082] Step S3: When the visual occlusion area of ​​the target cow exceeds a preset threshold, the pan / tilt of the adjacent camera group is dispatched to perform visual angle compensation and extract motion vector data of the adjacent cow bodies in the occlusion area;

[0083] Step S3 includes:

[0084] The PTZ is scheduled to perform horizontal rotation compensation and pitch angle adjustment. The horizontal rotation range is ±30° and the pitch angle adjustment is 10°-15°.

[0085] At the same time, the thermal imaging lens is activated to assist in positioning, and the temperature field change data is used to assist in spatial positioning;

[0086] Step S4, deducing the joint displacement dynamics of the occluded cow based on the motion vector data to generate a virtual skeleton action sequence;

[0087] Step S4 includes:

[0088] The spatiotemporal slicing attention mechanism is used to analyze the motion trajectories of adjacent cattle, and the skeletal displacement dynamics of the obscured cattle are deduced based on the spatial topological relationship of the herd.

[0089] The joint displacement dynamics include the change in angular velocity of the hind limb joints and the direction of the trunk center of gravity offset.

[0090] The dynamic deduction of joint displacement specifically includes:

[0091] a) Extract the displacement acceleration of adjacent bovine knee joints;

[0092] b) Calculate the force transfer coefficient of the blocked cattle based on the average distance between the cattle;

[0093] c) Generate the hind limb joint motion trajectory envelope by combining spatial topological constraints;

[0094] In the step of calculating the change in angular velocity of the hind limb joints in step S4, the coordinates of the hip, knee, and ankle points of the hind limbs are extracted from the three-dimensional skeleton model, and the following calculations are performed in sequence:

[0095] Calculate the joint angle:

[0096]

[0097] Among them, θ h (t) represents the angle of the hind limb joint at time t, v thigh (t) represents the vector from the hip joint to the knee joint, represented by p hip (t)-p knee (t) calculation, v shank (t) represents the vector from the knee joint to the ankle joint, represented by p ankle (t)-p knee (t) calculation, p hip (t) represents the spatial coordinate of the cow's hip joint at time t, p knee (t) represents the spatial coordinate of the cattle knee joint at time t, p ankle (t) represents the spatial coordinates of the ankle joint of the cow at time t, |·| represents the Euclidean norm of the vector;

[0098] Calculate the instantaneous angular velocity,

[0099]

[0100] Among them, ω h (t) represents the instantaneous angular velocity of the hind limb joint at time t, θ h (t-Δt) represents the joint angle at the previous frame time t-Δt, θ h In radians, ω h (t) is in radians per second, and Δt represents the time interval between two adjacent frame samples, in seconds;

[0101] Calculate the change in angular velocity:

[0102] Δω h (t) = ω h (t)-ω h (t-Δt),

[0103] Among them, Δω h (t) represents the change in angular velocity of the hind limb joints at time t, ω h (t-Δt) represents the instantaneous angular velocity of the previous frame;

[0104] Exponential smoothing filter:

[0105]

[0106] in, It represents the change in angular velocity after smoothing, α is the exponential smoothing coefficient, dimensionless, ranging from 0.2 to 0.3, and is used to balance response speed and noise suppression. Indicates the smoothed angular velocity change of the previous frame;

[0107] Smoothing the results As the hind limb dynamic features input to the virtual skeleton action sequence generation module;

[0108] Specifically, this method closely integrates with the three-dimensional skeletal model, accurately extracting joint motion changes through vector angles and time differences, then uses exponential smoothing to reduce the impact of video noise. Vector dot products and norms ensure the geometric accuracy of joint angle measurements, and instantaneous differences highlight the acceleration and deceleration trends of movements. The filtering process with adjustable smoothing coefficients effectively suppresses mutation errors while maintaining motion details. The entire process has a moderate amount of computation, suitable for real-time operation at edge nodes, and can provide stable and reliable angular velocity dynamic characteristics without the need for additional sensors.

[0109] Step S5: If the virtual skeleton motion sequence detects the associated motion of the hind limbs pushing off the ground and the neck extending forward, it is marked as a suspected climbing event;

[0110] Step S5 includes:

[0111] When the acceleration sensor returns the climbing characteristic signal, the fundamental frequency of the gait of the cattle in the previous period is synchronously analyzed;

[0112] If there is a limp walking feature, a dynamic suppression coefficient is applied to the suspected climbing event, and the suppression coefficient decreases as the duration of the limp feature increases;

[0113] Dynamic suppression coefficient generation includes:

[0114] When the duration of lameness is ≤5 minutes, the suppression coefficient is 0.1-0.3;

[0115] When the duration is greater than 5 minutes, the inhibition coefficient decreases by 0.05 per minute, with a lower limit of 0.6;

[0116] Step S6: fusing visible light texture and thermal imaging temperature field change data to perform multispectral verification on the suspected climbing event;

[0117] Step S6 includes:

[0118] Dynamically allocate the fusion weight of visible light and thermal imaging according to the ambient light intensity;

[0119] Under strong backlight conditions, the thermal imaging temperature gradient map is used to guide the edge restoration of visible light images;

[0120] Edge restoration uses a generative adversarial network, which takes the temperature field of thermal imaging as input to generate dust removal texture features.

[0121] The restored image is fed into the improved YOLOv8 model, which incorporates a channel attention mechanism in the feature extraction layer.

[0122] In step S6, after measuring the ambient light intensity at time t, a fusion weight calculation is performed, including:

[0123] Measure and normalize the ambient light, assuming:

[0124] I a (t) = S lux (t),

[0125] Among them, I a (t) represents the ambient light intensity at time t, unit: lux, S lux (t) represents the raw measurement value from the camera’s built-in light sensor;

[0126] Visible light weight map:

[0127]

[0128] in, Represents the uncropped visible light fusion weight, β represents the light response steepness coefficient, unit: lux -1 , I mIndicates the neutral light threshold, unit: lux,

[0129] In the formula, the steepness coefficient β and the neutral threshold I m Determined by the preset critical light range:

[0130]

[0131] Among them, W v,min Represents the minimum visible light weight, W v,max Indicates the highest visible light weight, L low Indicates the low light threshold, unit: lux, L high Indicates the high light threshold, unit: lux,

[0132] Before the system was deployed, ambient light intensity data was collected at different locations and time periods within the cowshed. Low and high light thresholds were determined based on the data distribution. These two thresholds served as dynamic distinction points, and when the real-time measured light intensity fell within this range, it was proportionally mapped to the visible light fusion weight range. To retain sufficient visible light detail even in low-light conditions, the lower limit of the visible light weight was fixed at 0.2; to avoid completely discarding thermal imaging information in extremely strong light conditions, the upper limit of the visible light weight was fixed at 0.8. During actual operation, the fusion ratio of visible light and thermal imaging was dynamically adjusted based on the relative position of the real-time light and these two thresholds, ensuring that the weights of both always remained within the preset range.

[0133]

[0134] Clip visible light weights:

[0135]

[0136] Among them, W v (t) represents the visible light fusion weight after clipping, max(·) and min(·) are the element-wise maximum and minimum operations, respectively;

[0137] Calculate the thermal imaging weight:

[0138] W t (t)=1-W v (t),

[0139] Among them, W t (t) represents the thermal imaging fusion weight,

[0140] Pixel-level image fusion:

[0141] I f (x,y,t)=W v (t)I vis (x,y,t)+W t(t)I th (x,y,t),

[0142] Among them, I f (x, y, t) represents the pixel intensity of the fused image, I vis (x, y, t) represents the visible light image pixel intensity, I th (x, y, t) represents the normalized pixel value of the thermal imaging image, and (x, y) represents the pixel coordinates;

[0143] The specific method for normalizing thermal imaging pixels is as follows: based on the range calibration of the thermal imaging camera, the original temperature value of each pixel in each frame of the image is subtracted from the lower limit of the sensor range, and then divided by the difference between the upper and lower limits of the sensor range, and the resulting value is mapped to the range of 0 to 1; so that all thermal imaging pixels are on the same normalized scale, which facilitates pixel-level fusion with visible light images with the same weight.

[0144] Specifically, this method is based on real-time light measurement and dynamically converts ambient light into fusion weights for visible light and thermal imaging through a logistic mapping function with a threshold interval. The steepness of the mapping function curve is automatically determined by the parameters within the critical light interval, without the need for manual parameter adjustment. The clipping step ensures that the weights always fall within the preset minimum and maximum ranges, avoiding extreme light conditions that cause the weights to be excessively biased towards a single modality. The final pixel-level fusion can not only fully utilize the advantages of thermal imaging in strong backlight or dim environments, but also take advantage of the high resolution of visible light when there is sufficient light. This solution has a simple algorithm and moderate computational complexity, making it suitable for real-time deployment at edge nodes and effectively improving the quality of multispectral images.

[0145] Step S7, outputting the estrus behavior determination result to the ranch management system;

[0146] The edge computing node adopts a regional processing architecture, with the camera's built-in FPGA performing target detection and the edge server running a cross-lens spatiotemporal alignment algorithm;

[0147] The spatiotemporal alignment algorithm uses a 3D skeleton model database as a benchmark to map multiple video streams into a unified coordinate system;

[0148] The cross-shot spatiotemporal alignment algorithm includes a skeletal key point matching module, which realizes multi-view data association through the spatial position of the cow's scapula and hip joint.

[0149] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A cattle estrus monitoring method based on multi-camera data fusion, characterized in that: include, Step S1: deploy at least four wide-angle camera groups on the ceiling of the cowshed, each group including a visible light lens and a thermal imaging lens, and each lens is connected to the edge computing node through a servo pan / tilt; Step S2: The edge computing node acquires multiple video streams in real time and constructs a three-dimensional skeletal model database of cattle based on the cattle's back pattern and body contour features; Step S3: When the visual occlusion area of ​​the target cow exceeds a preset threshold, the pan / tilt of the adjacent camera group is dispatched to perform visual angle compensation and extract motion vector data of the adjacent cow bodies in the occlusion area; Step S4, deducing the joint displacement dynamics of the occluded cow based on the motion vector data to generate a virtual skeleton action sequence; Step S5: If the virtual skeleton motion sequence detects the associated motion of the hind limbs pushing off the ground and the neck extending forward, it is marked as a suspected climbing event; Step S6: fusing visible light texture and thermal imaging temperature field change data to perform multispectral verification on the suspected climbing event; Step S7: outputting the estrus behavior determination result to the pasture management system.

2. The cattle estrus monitoring method based on multi-camera data fusion according to claim 1, characterized in that: Step S3 includes: The PTZ is scheduled to perform horizontal rotation compensation and pitch angle adjustment. The horizontal rotation range is ±30° and the pitch angle adjustment is 10°-15°. At the same time, the thermal imaging lens is activated to assist in positioning, and the temperature field change data is used to assist in spatial positioning.

3. The cattle estrus monitoring method based on multi-camera data fusion according to claim 1, characterized in that: Step S4 includes: The spatiotemporal slicing attention mechanism is used to analyze the motion trajectories of adjacent cattle, and the skeletal displacement dynamics of the obscured cattle are deduced based on the spatial topological relationship of the herd. The joint displacement dynamics include the change in angular velocity of the hind limb joints and the direction of the trunk center of gravity offset.

4. A cattle estrus monitoring method based on multi-camera data fusion as claimed in claim 3, characterized in that: The joint displacement dynamic deduction specifically includes: a) Extract the displacement acceleration of adjacent bovine knee joints; b) Calculate the force transfer coefficient of the blocked cattle based on the average distance between the cattle; c) Generate the envelope of the hind limb joint motion trajectory by combining spatial topological constraints.

5. The cattle estrus monitoring method based on multi-camera data fusion according to claim 4, characterized in that: In the step of calculating the change in angular velocity of the hind limb joints in step S4, the coordinates of the hind limb hip-knee-ankle points are extracted from the three-dimensional skeleton model, and the following calculations are performed in sequence: Calculate the joint angle: Among them, θ h (t) represents the angle of the hind limb joint at time t, v thigh (t) represents the vector from the hip joint to the knee joint, represented by p hip (t)-p knee (t) calculation, v shank (t) represents the vector from the knee joint to the ankle joint, represented by p ankle (t)-p knee (t) calculation, p hip (t) represents the spatial coordinate of the cow's hip joint at time t, p knee (t) represents the spatial coordinate of the cattle knee joint at time t, p ankle (t) represents the spatial coordinates of the ankle joint of the cow at time t, |·| represents the Euclidean norm of the vector; Calculate the instantaneous angular velocity, Among them, ω h (t) represents the instantaneous angular velocity of the hind limb joint at time t, θ h (t-Δt) represents the joint angle at the previous frame time t-Δt, θ h In radians, ω h (t) is in radians per second, and Δt represents the time interval between two adjacent frame samples, in seconds; Calculate the change in angular velocity: Give h (t)=ω h (t)-ω h (t-Δt), Among them, Δω h (t) represents the change in angular velocity of the hind limb joints at time t, ω h (t-Δt) represents the instantaneous angular velocity of the previous frame; Exponential smoothing filter: in, It represents the change in angular velocity after smoothing, α is the exponential smoothing coefficient, dimensionless, ranging from 0.2 to 0.3, and is used to balance response speed and noise suppression. Indicates the smoothed angular velocity change of the previous frame; Smoothing the results It is used as the dynamic feature of the hind limbs and is input into the virtual skeleton action sequence generation module.

6. The cattle estrus monitoring method based on multi-camera data fusion according to claim 1, characterized in that: Step S5 includes: When the acceleration sensor returns the climbing characteristic signal, the fundamental frequency of the gait of the cattle in the previous period is synchronously analyzed; If limping walking characteristics are present, a dynamic suppression coefficient is applied to the suspected climbing event, and the suppression coefficient decreases as the duration of the limping characteristics increases.

7. The cattle estrus monitoring method based on multi-camera data fusion according to claim 1, characterized in that: Step S6 includes: Dynamically allocate the fusion weight of visible light and thermal imaging according to the ambient light intensity; Under strong backlight conditions, the thermal imaging temperature gradient map is used to guide the edge restoration of visible light images; The edge restoration adopts a generative adversarial network, which takes the temperature field of thermal imaging as a conditional input to generate dust removal texture features; The repaired image is input into the improved YOLOv8 model, which incorporates a channel attention mechanism in the feature extraction layer.

8. The cattle estrus monitoring method based on multi-camera data fusion according to claim 7, characterized in that: In step S6, after measuring the ambient light intensity at time t, a fusion weight calculation is performed, including: Measure and normalize the ambient light, assuming: I a (t)=S lux (t), Among them, I a (t) represents the ambient light intensity at time t, unit: lux, S lux (t) represents the raw measurement value from the camera’s built-in light sensor; Visible light weight map: in, Represents the uncropped visible light fusion weight, β represents the light response steepness coefficient, unit: lux -1 , I m Indicates the neutral light threshold, unit: lux, In the formula, the steepness coefficient β and the neutral threshold I m Determined by the preset critical light range: Among them, W v,min Represents the minimum visible light weight, W v,max Indicates the highest visible light weight, L low Indicates the low light threshold, unit: lux, L high Indicates the high light threshold, unit: lux, Clip visible light weights: Among them, W v (t) represents the visible light fusion weight after clipping, max(·) and min(·) are the element-wise maximum and minimum operations, respectively; Calculate the thermal imaging weight: W t (t)=1-W v (t), Among them, W t (t) represents the thermal imaging fusion weight, Pixel-level image fusion: I f (x,y,t)=W v (t)I vis (x,y,t)+W t (t)I th (x,y,t), Among them, I f (x, y, t) represents the pixel intensity of the fused image, I vis (x, y, t) represents the visible light image pixel intensity, I th (x, y, t) represents the normalized pixel value of the thermal imaging image, and (x, y) represents the pixel coordinates.

9. The cattle estrus monitoring method based on multi-camera data fusion according to claim 1, characterized in that: The edge computing node adopts a regional processing architecture, the camera's built-in FPGA performs target detection, and the edge server runs a cross-lens spatiotemporal alignment algorithm; The spatiotemporal alignment algorithm uses the three-dimensional skeleton model database as a benchmark to map multiple video streams into a unified coordinate system.

10. The cattle estrus monitoring method based on multi-camera data fusion according to claim 9, characterized in that: The cross-shot spatiotemporal alignment algorithm includes a skeletal key point matching module, which realizes multi-view data association through the spatial positions of the cow's shoulder blade and hip joint.

Images