Method for monitoring health state of sheep through gait recognition

Abstract

The invention relates to the technical field of sheep monitoring, in particular to a method for monitoring the health state of sheep through gait recognition, which comprises the following steps: acquiring a sheep walking video stream through a camera, reconstructing and generating a three-dimensional skeleton motion sequence based on a skeleton motion track, and marking space-time coordinates of hip joint, knee joint and hoof key points; according to the three-dimensional skeleton motion sequence, the knee joint buckling angular velocity peak value, the same-side front and rear hoof grounding time difference and the trunk lateral swing amplitude in a single-step period are calculated, and a gait space-time coordination parameter set is constructed; and inputting the gait time-space coordination parameter set into a disease decoupling model based on anti-factual reasoning, and outputting an early arthritis risk index and a neurological dyskinesia probability. According to the invention, the arthritis risk index and the neurological dyskinesia probability are calculated, early recognition and risk grading early warning of dyskinesia are realized, and the intelligent level and the response capability of a health monitoring system are remarkably improved.

Description

Technical Field

[0001] The present invention relates to the technical field of sheep monitoring, and in particular to a method for monitoring the health status of sheep by utilizing gait recognition. Background Art

[0002] As important herbivorous livestock, the health of sheep directly affects the economic benefits and disease prevention of the animal husbandry industry. Traditional health monitoring methods rely heavily on manual observation and regular physical examinations, which suffer from low detection frequency, large subjective errors, and delayed response. This is especially difficult to identify in the early stages of behavioral abnormalities such as arthritis and neurological movement disorders, when they are not yet obvious. Furthermore, existing gait analysis systems based on a single image modality struggle to accurately extract skeletal motion information under complex lighting conditions, hair occlusion, and ground interference, resulting in insufficient accuracy in gait parameter extraction.

[0003] In recent years, 3D reconstruction and deep learning technologies have been gradually applied in animal behavior analysis. However, these methods primarily focus on static posture recognition or gait classification, lacking interpretable indicators for specific pathological conditions. Furthermore, existing methods generally fail to account for individual differences and environmental factors, failing to effectively decouple the causes of gait abnormalities. This in turn limits the diagnostic and risk warning capabilities of intelligent monitoring systems. Summary of the Invention

[0004] The present invention provides a method for monitoring the health status of sheep using gait recognition. This health status monitoring method integrates multimodal perception, three-dimensional skeleton reconstruction, gait feature extraction and counterfactual reasoning capabilities. It can not only characterize the movement characteristics of individual sheep, but also separate pathological and environmental influencing factors, realizing intelligent diagnosis and dynamic early warning for early disease risks.

[0005] A method for monitoring the health status of sheep using gait recognition comprises the following steps:

[0006] S1. Capture the sheep's walking video stream through a camera, reconstruct a 3D skeletal motion sequence based on the skeletal motion trajectory, and mark the spatiotemporal coordinates of key points of the hip joint, knee joint, and hoof;

[0007] S2. Calculate the peak knee flexion angular velocity, the ground contact time difference between the front and rear hooves on the same side, and the lateral swing amplitude of the trunk within a single step cycle based on the three-dimensional skeletal motion sequence, and construct a gait spatiotemporal coordination parameter set;

[0008] S3. Input the gait spatiotemporal coordination parameter set into a pathological decoupling model based on counterfactual reasoning, and output an early arthritis risk index and a probability of neurological movement disorders by comparing the gait phase-mechanical coupling relationship of sheep of the same breed in a healthy benchmark library.

[0009] Optionally, the S1 specifically includes:

[0010] S11: Visible light and near-infrared dual-band cameras are symmetrically deployed on both sides of the sheep house passage, and a short-wave infrared camera is deployed on the top to form a multi-spectral stereo vision acquisition system to synchronously obtain video streams in predetermined bands;

[0011] S12, using improved Graph-CNN network to process multispectral video streams:

[0012] a) Through cross-modal feature fusion, visible light texture information and near-infrared skeleton contour features are fused with channel attention weighted to generate a high-confidence keypoint heat map;

[0013] b) constructing a graph structure model based on the sheep's anatomical topological constraints, wherein the graph structure model has the hip joint as the root node and the knee joint and hoof as child nodes, and iteratively optimizing the spatial coordinates of the joint points through graph convolution;

[0014] S13, input the multi-view video stream into the motion structure recovery algorithm, solve the three-dimensional space-time coordinates of the key points of the hip joint, knee joint and hoof, arrange the three-dimensional space-time coordinates in the time sequence of the video frames, form a coordinate set of each key point at each time step, and constitute a three-dimensional skeletal motion sequence of continuous frames.

[0015] Optionally, the graph structure model is represented as: G = (V, E), where the vertex set includes the hip joint, knee joint, and hoof, the edge set E represents the anatomical connection between the nodes, and the key point spatial position p i =(x i ,y i ,z i ) Iterative optimization through graph convolution:

[0016]

[0017] in, represents the set of adjacent nodes of node i, c ij represents the normalization coefficient between nodes i and j; W (l) represents the graph convolution weight matrix of the lth layer, δ is the nonlinear activation function, The coordinate features of node i in the lth layer are used to constrain the key bone length to be optimized through the preset structural constraints.

[0018] Optionally, in the three-dimensional space-time coordinates: the x-axis and y-axis coordinates are obtained by stereoscopic visual disparity calculation, the z-axis coordinate is calibrated according to the geometric projection constraint at the moment the hoof touches the ground, and the timestamp is aligned with the video frame rate.

[0019] Optionally, the S2 specifically includes:

[0020] S21 Calculation of Knee Flexion Angular Velocity Peak: Based on the three-dimensional coordinates of the hip and knee joints, the knee flexion angle timing curve is reconstructed using the quaternion interpolation method. The angular velocity curve is obtained by first-order differentiation of the angle curve, and the maximum angular velocity value and the corresponding swing phase within a single step cycle are extracted.

[0021] S22, calculation of the time difference between the front and rear hoof contact on the same side: using the moment of front hoof contact as the reference point, detect the moment when the z-axis coordinate of the key point of the hoof first falls below the ground height threshold and the speed direction is from bottom to top, align the timestamps of the front and rear hoof contact events on the same side using the dynamic time warping algorithm, and calculate the time difference between the two;

[0022] S23, calculation of trunk lateral swing amplitude: using the line connecting the midpoints of the bilateral hip joints as the trunk reference axis, calculate the deflection angle of this axis from the vertical direction in each frame image, and calculate the standard deviation of the deflection angle within a single step cycle as a quantitative indicator of the lateral swing amplitude;

[0023] S24, normalizing the angular velocity peak value, ground contact time difference, and swing amplitude according to the gait phase, combining them into a multidimensional feature vector, and adding a ground slope compensation factor to form a gait spatiotemporal coordination parameter set.

[0024] Optionally, the calculation of the contact time difference between the front and rear hooves on the same side in S22 specifically includes:

[0025] The moment of ground contact is determined based on the change in the vertical height of the hoof and the speed of movement. When the hoof height is lower than the ground height threshold and the speed direction changes from downward to upward, it is determined to be a ground contact event.

[0026] The contact times of the front and hind hooves were recorded separately, and their time intervals were calculated. The two sets of contact time series were aligned using a dynamic time warping algorithm to correct the timing error caused by gait asymmetry. The time difference was compensated according to the slope angle of the walking path to obtain a true synchronization index.

[0027] Optionally, the calculation of the lateral swing amplitude of the trunk in S23 specifically includes:

[0028] A trunk reference axis is constructed using the line connecting the two hip joints, and the degree of deflection of the trunk reference axis relative to the vertical direction in each frame of the image is determined to obtain a continuous deflection angle sequence;

[0029] In a single complete gait cycle, the standard deviation of the deflection angle sequence is calculated to measure the lateral swing stability of the trunk, reflecting the changes in individual motor control coordination.

[0030] Optionally, the S3 specifically includes:

[0031] S31, Counterfactual Sample Construction: Generative Adversarial Networks are used to simulate the spatiotemporal coordination parameter sets of the target sheep's gait under ideal ground conditions and without pathological interference to generate counterfactual health benchmark sequences;

[0032] S32, Timing Difference Calculation: Align the gait phase of the actual parameter set with the counterfactual healthy baseline sequence using a dynamic time warping algorithm; calculate the peak difference of the knee flexion angular velocity, the ground contact time difference offset, and the trunk swing amplitude residual to generate a multidimensional difference vector;

[0033] S33, causal effect decomposition: Construct a structural equation model to decompose the difference vector into pathological contribution factors and environmental interference factors. Quantify the weight of the pathological contribution factor in the difference vector through maximum likelihood estimation. If the pathological contribution factor in the difference vector has a pathologically dominant weight greater than the preset weight threshold, the pathological factor is determined to be dominant.

[0034] S34, Disease Risk Assessment:

[0035] The arthritis risk index is calculated based on the product of the difference in knee flexion angular velocity peak value and the pathology weight ratio. The risk level is divided and a graded warning is triggered.

[0036] The probability of neurological movement disorder is outputted by a joint probability distribution model based on the contact time difference offset and the trunk swing residual, and the probability value is mapped to the range of 0-100%.

[0037] Optionally, in the causal effect decomposition, the pathological contribution factor is associated with the mechanical response pattern of joint inflammation and nerve damage; and the environmental interference factor is associated with the residual noise after ground slope compensation.

[0038] Optionally, the generative adversarial network includes a generator G(·) and a discriminator D(·), wherein:

[0039] Generator G(·) outputs the counterfactual health benchmark sequence G cf , which satisfies: G cf =G(G real ,meta),G real It represents the spatiotemporal coordination parameter set of the target sheep’s gait currently observed, and meta represents the individual characteristic information of the sheep, including breed, weight, and age data;

[0040] The discriminator D(·) optimizes the generation quality by comparing the generated sequences with the distribution in the healthy benchmark library.

[0041] Beneficial effects of the present invention:

[0042] This invention builds a multispectral stereoscopic vision acquisition system that integrates visible light, near-infrared, and short-wave infrared light, and introduces a graph convolutional network with a channel attention mechanism and anatomical topology constraint optimization to accurately extract the three-dimensional coordinates of key points of the hip, knee, and hoof. Through stereo disparity calculation and geometric projection calibration of the touchdown point, the accuracy of positioning along the x, y, and z axes is improved, adapting to day and night environments and complex ground conditions, effectively supporting subsequent health status analysis models.

[0043] The present invention designs a unified gait spatiotemporal feature calculation process based on the knee flexion angular velocity, the ipsilateral hoof contact time difference and the lateral swing amplitude of the trunk, and improves stability and comparability through quaternion interpolation, Kalman filtering and gait phase standardization. At the same time, a slope angle correction mechanism is introduced to eliminate the influence of environmental interference on parameter calculation, making the generated parameter set more diagnostic and discriminative, and providing a clear quantitative basis for disease risk determination.

[0044] The present invention uses a GAN model to generate a healthy parameter sequence without pathological interference as a control, and combines the dynamic time warping alignment strategy with the structural equation model to decompose the source of differences. It can decompose the observed gait abnormality into two parts: pathological contribution and environmental disturbance. Based on maximum likelihood estimation, the pathological weight is accurately evaluated, and the arthritis risk index and the probability of neurological movement disorders are further calculated. This realizes early identification of movement disorders and risk classification warning, significantly improving the intelligence level and responsiveness of the health monitoring system. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only for the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0046] Figure 1 Schematic diagram of a method flow in an embodiment of the present invention;

[0047] Figure 2 Schematic diagram of constructing a gait spatiotemporal coordination parameter set according to an embodiment of the present invention. DETAILED DESCRIPTION

[0048] The present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. Those skilled in the art may also implement some known technologies in other alternative ways. The accompanying drawings are only for describing the embodiments in more detail and are not intended to limit the present invention in any specific way.

[0049] like Figure 1-Figure 2 As shown, a method for monitoring the health status of sheep using gait recognition includes the following steps:

[0050] S1. Capture the sheep's walking video stream through a camera, reconstruct a 3D skeletal motion sequence based on the skeletal motion trajectory, and mark the spatiotemporal coordinates of key points of the hip joint, knee joint, and hoof;

[0051] S2. Based on the three-dimensional skeletal motion sequence, the peak knee flexion angular velocity, the time difference between the front and rear hooves on the same side and the lateral swing amplitude of the trunk in a single step cycle are calculated to construct a gait spatiotemporal coordination parameter set;

[0052] S3. Input the gait spatiotemporal coordination parameter set into the pathological decoupling model based on counterfactual reasoning, and output the early arthritis risk index and the probability of neurological movement disorders by comparing the gait phase-mechanical coupling relationship of sheep of the same breed in the health benchmark library.

[0053] S1 specifically includes:

[0054] Visible light and near-infrared dual-band cameras are deployed symmetrically on both sides of the sheep house channel, and a short-wave infrared camera is deployed on the top to build a multi-spectral stereo vision acquisition system to synchronously obtain video stream data with a wavelength range of 500-1700nm.

[0055] For the collected multi-view and multi-spectral video streams, an improved graph convolutional neural network (Graph-CNN) is used to extract key points and perform skeleton topology modeling. The steps are as follows:

[0056] (a) Cross-modal feature fusion and heat map generation:

[0057] The visible light texture feature map F is obtained by channel attention mechanism vis and near-infrared bone contour feature map F nir Perform weighted fusion to generate key point heat map H k :

[0058] H k =σ(W·[α vis ·F vis ||α nir ·F nir ]);

[0059] Among them, α vis , α nir is the channel attention weight, || represents the feature channel splicing, W is the fused convolution kernel, σ is the activation function, H k is the heat map of the kth key point.

[0060] (b) Dynamic graph structure skeleton modeling and coordinate optimization: Construct a dynamic skeleton graph G = (V, E) with topological constraints, where the vertex set V includes the hip joint (root node), knee joint, and hoof (child nodes), and the edge set E represents the anatomical connection between each node. Key point spatial position pi =(x i ,y i ,z i ) Iterative optimization through graph convolution:

[0061]

[0062] in, represents the set of adjacent nodes of node i, c ij represents the normalization coefficient between nodes i and j; W (l) represents the graph convolution weight matrix of the lth layer, δ is the nonlinear activation function, is the coordinate feature of node i in layer l.

[0063] Key bone lengths are optimized using pre-set structural constraints, for example, hip span is approximately 15 cm and tibia length is approximately 20-25 cm.

[0064] (c) Three-dimensional coordinate solution and time synchronization processing:

[0065] 1. XY coordinate estimation: obtained by calculating the stereo disparity of the left and right cameras:

[0066] Where, B is the baseline distance (recommended value 1.2m), d x ,d y is the pixel parallax between the left and right cameras, and f is the focal length of the camera.

[0067] 2. Z-coordinate estimation: Based on the top short-wave infrared camera, when the hoof touches the ground (the thermal conductivity transition point) and its vertical velocity is set to zero, the z-axis geometric constraint equation is constructed by projection:

[0068] This condition is used for z-axis coordinate correction, t TD is the hoof contact time, Δt is the time sampling interval, v z is the speed of the hoof in the z-axis direction.

[0069] 3. Timestamp alignment: Each camera uses PTP to achieve frame-level synchronization, ensuring that key point trajectories are consistent in time under multiple viewing angles.

[0070] S2 specifically includes:

[0071] S21, calculation of peak knee flexion angular velocity:

[0072] (a) Angle sequence reconstruction (based on quaternion interpolation): hip joint coordinates p h (t)=(x h (t),y h (t),z h(t)) (three-dimensional coordinates of the hip joint at time t) and the knee joint coordinates p k (t)=(x k (t), y k (t), z k (t)) (the three-dimensional coordinates of the knee joint at time t) are used as input to construct a unit direction vector sequence:

[0073] Will Convert it into a rotation quaternion trajectory q(t), and reconstruct the continuous flexion angle sequence θ(t) through SLERP (spherical linear interpolation). represents the unit direction vector from the hip joint to the knee joint, and θ(t) represents the curve of the knee flexion angle changing with time.

[0074] (b) Angular velocity calculation and peak extraction Perform first-order time differentiation on θ(t) to obtain the knee joint angular velocity curve: Extract single-step cycle [t s ,t e ] maximum value within: t s ,t e is the start and end time point of the single-step cycle, ω max is the peak value of knee flexion angular velocity during a single step cycle.

[0075] S22, calculation of the difference in time between the front and rear hoof contact on the same side:

[0076] (a) Touchdown event detection: Let the three-dimensional coordinates of the key point of the front hoof be p f (t), the three-dimensional coordinates of the key point of the hind hoof are p b (t), the touchdown moment is defined as:

[0077] t TD =min{t|z(t) <z th ,v z (t)>0}; where z th Indicates the ground height threshold, determined by calibration, represents the z-axis velocity of the hoof, z(t) represents the vertical height of the key point, and the time difference between the front hoof and the hind hoof is estimated by the Kalman filter and is defined as: Indicates the moment when the front and rear hooves touch the ground (the ground contact conditions are met);

[0078] (b) Time difference alignment and correction: The dynamic time warping algorithm is used to align the trajectory sequences of the front and rear hoof contact events, eliminating asynchronous interference and considering the walking slope angle α. The final time difference is:

[0079] α is the slope angle of the walking path, Indicates the time difference after slope compensation is taken into account.

[0080] S23, calculation of trunk lateral swing amplitude:

[0081] (a) Construction of the trunk reference axis:

[0082] The midpoints of the bilateral hip joints are: p c (t) represents the three-dimensional coordinates of the midpoints of the bilateral hip joints at time t, x hL (t) represents the x coordinate of the left hip joint at time t, x hR (t) represents the x coordinate of the right hip joint at time t, x hL (t) represents the y coordinate of the left hip joint at time t, y hR (t) represents the y coordinate of the right hip joint at time t, z hL (t) represents the z coordinate of the left hip joint at time t, z hR (t) represents the z coordinate of the right hip joint at time t;

[0083] Construct the torso vector: represents the trunk axis vector (trunk reference axis) formed by the line connecting the two hip joints at time t;

[0084] Let the vertical unit vector be The deflection angle of each frame is:

[0085] θ(t) represents the angle (deflection angle) between the trunk axis and the vertical direction in each frame image;

[0086] (b) Quantification of the swing amplitude: Statistical analysis of the standard deviation σ of the deflection angle within a single step period θ :

[0087] Represents the average value of the deflection angle within a single step cycle, N θ Indicates the number of frames in the gait cycle.

[0088] S24, construction of gait spatiotemporal coordination parameter set: the peak angular velocity ω max , time difference after correction and the standard deviation of the swing amplitude σ θ Perform normalization (denoted as ), construct the gait spatiotemporal coordination parameter vector:

[0089] Based on the pathology decoupling model of counterfactual reasoning, a counterfactual healthy control sample is constructed for the target sheep's gait spatiotemporal coordination parameter set, and the pathological influence and environmental interference are separated to finally output the disease risk index. Therefore, S3 specifically includes:

[0090] S31, counterfactual sample construction: Generative Adversarial Network (GAN) is used to simulate the healthy gait parameters of the target sheep under ideal conditions:

[0091] Input features include sheep breed, weight, age and actual gait spatiotemporal coordination parameter set G real ;

[0092] Generator G(·) outputs the counterfactual health benchmark sequence G cf , which satisfies: G cf =G(G real ,meta);

[0093] The discriminator D(·) optimizes the generation quality by comparing the generated sequences with the distribution in the healthy benchmark library.

[0094] S32, timing difference calculation:

[0095] The gait phase alignment and difference quantification of the actual parameter sequence and the counterfactual health sequence were performed, and the dynamic time warping alignment method was applied to G real With G cf Perform phase synchronization to ensure that key events correspond on the timeline;

[0096] Difference vector construction: ΔG = [ΔV peak ,Δt touch ,θS sway ], where ΔV peak =V real -V cf Indicates the difference in knee flexion angular velocity peak value, Δt touch =t real -t cf Indicates the contact time difference offset iΔS sway =σ real -σ cf Represents the residual error of trunk swing amplitude.

[0097] S33, causal effect decomposition: Based on the structural equation model (SEM), the difference vector is decomposed into two types of contributions: pathological factors and environmental factors: ΔG = λ p ·F patho +λ e ·F env +∈; where F patho represents the pathological contribution factor (arthritis, nerve damage), F envrepresents the environmental interference factor (ground roughness, slope residue), λ p ,λ e represents the path coefficient (weight), and ∈ is the Gaussian noise term (residual error).

[0098] Optimize via maximum likelihood estimation: like It is determined to be dominated by pathological factors.

[0099] S34, Disease Risk Assessment:

[0100] S341, Arthritis risk index calculation: R arthritis =ΔV peak w p +|Δt stance |·w t ; Among them, w p Pathological factor weight (derived from SEM), w t is the weight of the time difference of the support period, Δt stance is the support period offset.

[0101] The risk levels are as follows:

[0102] R<30: low risk;

[0103] 30≤R<60: medium risk;

[0104] R≥60: High risk, triggering an early warning.

[0105] S342, Probability Assessment of Neurological Movement Disorders:

[0106] Among them, N c represents the nerve damage event, Δt touch ,ΔS swsy are the touchdown time difference offset and the swing amplitude residual, respectively, and P (Evidence) represents the marginal probability of the joint observation event. neuro When the value is >0.75, further neurological examination is recommended.

[0107] The present invention encompasses any alternatives, modifications, equivalents, and solutions that fall within the spirit and scope of the present invention. To provide a thorough understanding of the present invention, specific details are described in detail below in connection with the preferred embodiments of the present invention, but those skilled in the art will be able to fully understand the present invention without these detailed descriptions. Furthermore, to avoid unnecessary confusion regarding the essence of the present invention, well-known methods, processes, procedures, components, and circuits have not been described in detail.

[0108] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications should also be regarded as within the scope of protection of the present invention.

Claims

1. A method for monitoring the health status of sheep using gait recognition, characterized in that: The following steps are involved: S1. Capture the sheep's walking video stream through a camera, reconstruct a 3D skeletal motion sequence based on the skeletal motion trajectory, and mark the spatiotemporal coordinates of key points of the hip joint, knee joint, and hoof; S2. Calculate the peak knee flexion angular velocity, the ground contact time difference between the front and rear hooves on the same side, and the lateral swing amplitude of the trunk within a single step cycle based on the three-dimensional skeletal motion sequence, and construct a gait spatiotemporal coordination parameter set; S3. Input the gait spatiotemporal coordination parameter set into a pathological decoupling model based on counterfactual reasoning, and output an early arthritis risk index and a probability of neurological movement disorders by comparing the gait phase-mechanical coupling relationship of sheep of the same breed in a healthy benchmark library.

2. The method for monitoring the health status of sheep using gait recognition according to claim 1, characterized in that: Said S1 specifically includes: S11: Visible light and near-infrared dual-band cameras are symmetrically deployed on both sides of the sheep house passage, and a short-wave infrared camera is deployed on the top to form a multi-spectral stereo vision acquisition system to synchronously obtain video streams in predetermined bands; S12, using improved Graph-CNN network to process multispectral video streams: a) Through cross-modal feature fusion, visible light texture information and near-infrared skeleton contour features are fused with channel attention weighted to generate a high-confidence keypoint heat map; b) constructing a graph structure model based on the sheep's anatomical topological constraints, wherein the graph structure model has the hip joint as the root node and the knee joint and hoof as child nodes, and iteratively optimizing the spatial coordinates of the joint points through graph convolution; S13, input the multi-view video stream into the motion structure recovery algorithm, solve the three-dimensional space-time coordinates of the key points of the hip joint, knee joint and hoof, arrange the three-dimensional space-time coordinates in the time sequence of the video frames, form a coordinate set of each key point at each time step, and constitute a three-dimensional skeletal motion sequence of continuous frames.

3. The method for monitoring the health status of sheep using gait recognition according to claim 2, characterized in that: The graph structure model is represented as: G = (V, E), where the vertex set includes the hip joint, knee joint, and hoof, the edge set E represents the anatomical connection between the nodes, and the key point spatial position p i =(x i ,y i ,z i ) Iterative optimization through graph convolution: in, represents the set of adjacent nodes of node i, c ij represents the normalization coefficient between nodes i and j; W (l) represents the graph convolution weight matrix of the lth layer, δ is the nonlinear activation function, The coordinate features of node i in the lth layer are used to constrain the key bone length to be optimized through the preset structural constraints.

4. The method for monitoring the health status of sheep using gait recognition according to claim 2, characterized in that: In the three-dimensional space-time coordinates: the x-axis and y-axis coordinates are obtained by stereoscopic visual parallax calculation, the z-axis coordinate is calibrated according to the geometric projection constraint at the moment the hoof touches the ground, and the timestamp is aligned with the video frame rate.

5. The method for monitoring the health status of sheep using gait recognition according to claim 1, characterized in that: The S2 specifically includes: S21 Calculation of Knee Flexion Angular Velocity Peak: Based on the three-dimensional coordinates of the hip and knee joints, the knee flexion angle timing curve is reconstructed using the quaternion interpolation method. The angular velocity curve is obtained by first-order differentiation of the angle curve, and the maximum angular velocity value and the corresponding swing phase within a single step cycle are extracted. S22, calculation of the time difference between the front and rear hoof contact on the same side: using the moment of front hoof contact as the reference point, detect the moment when the z-axis coordinate of the key point of the hoof first falls below the ground height threshold and the speed direction is from bottom to top, align the timestamps of the front and rear hoof contact events on the same side using the dynamic time warping algorithm, and calculate the time difference between the two; S23, calculation of trunk lateral swing amplitude: using the line connecting the midpoints of the bilateral hip joints as the trunk reference axis, calculate the deflection angle of this axis from the vertical direction in each frame image, and calculate the standard deviation of the deflection angle within a single step cycle as a quantitative indicator of the lateral swing amplitude; S24, normalizing the angular velocity peak value, ground contact time difference, and swing amplitude according to the gait phase, combining them into a multidimensional feature vector, and adding a ground slope compensation factor to form a gait spatiotemporal coordination parameter set.

6. The method for monitoring the health status of sheep using gait recognition according to claim 5, characterized in that: The calculation of the contact time difference between the front and rear hooves on the same side in S22 specifically includes: The moment of ground contact is determined based on the change in the vertical height of the hoof and the speed of movement. When the hoof height is lower than the ground height threshold and the speed direction changes from downward to upward, it is determined to be a ground contact event. The contact times of the front and hind hooves were recorded separately, and their time intervals were calculated. The two sets of contact time series were aligned using a dynamic time warping algorithm to correct the timing error caused by gait asymmetry. The time difference was compensated according to the slope angle of the walking path to obtain a true synchronization index.

7. The method for monitoring the health status of sheep using gait recognition according to claim 5, characterized in that: The calculation of the lateral swing amplitude of the trunk in S23 specifically includes: A trunk reference axis is constructed using the line connecting the two hip joints, and the degree of deflection of the trunk reference axis relative to the vertical direction in each frame of the image is determined to obtain a continuous deflection angle sequence; In a single complete gait cycle, the standard deviation of the deflection angle sequence is calculated to measure the lateral swing stability of the trunk, reflecting the changes in individual motor control coordination.

8. The method for monitoring the health status of sheep using gait recognition according to claim 1, characterized in that: The S3 specifically includes: S31, Counterfactual Sample Construction: Generative Adversarial Networks are used to simulate the spatiotemporal coordination parameter sets of the target sheep's gait under ideal ground conditions and without pathological interference to generate counterfactual health benchmark sequences; S32, Timing Difference Calculation: Align the gait phase of the actual parameter set with the counterfactual healthy baseline sequence using a dynamic time warping algorithm; calculate the peak difference of the knee flexion angular velocity, the ground contact time difference offset, and the trunk swing amplitude residual to generate a multidimensional difference vector; S33, causal effect decomposition: Construct a structural equation model to decompose the difference vector into pathological contribution factors and environmental interference factors. Quantify the weight of the pathological contribution factor in the difference vector through maximum likelihood estimation. If the pathological contribution factor in the difference vector has a pathologically dominant weight greater than the preset weight threshold, the pathological factor is determined to be dominant. S34, Disease Risk Assessment: The arthritis risk index is calculated based on the product of the difference in knee flexion angular velocity peak value and the pathology weight ratio. The risk level is divided and a graded warning is triggered. The probability of neurological movement disorder is outputted by a joint probability distribution model based on the contact time difference offset and the trunk swing residual, and the probability value is mapped to the range of 0-100%.

9. The method for monitoring the health status of sheep using gait recognition according to claim 8, characterized in that: In the causal effect decomposition, the pathological contribution factor is associated with the mechanical response pattern of joint inflammation and nerve damage; the environmental interference factor is associated with the residual noise after ground slope compensation.

10. The method for monitoring the health status of sheep using gait recognition according to claim 8, characterized in that: The generative adversarial network includes a generator G(·) and a discriminator D(·), where: Generator G(·) outputs the counterfactual health benchmark sequence G cf , which satisfies: G cf =G(G real ,meta),G real It represents the spatiotemporal coordination parameter set of the target sheep’s gait currently observed, and meta represents the individual characteristic information of the sheep, including breed, weight, and age data; The discriminator D(·) optimizes the generation quality by comparing the generated sequences with the distribution in the healthy benchmark library.

Images