A beef cattle abnormal behavior grading recognition method and system

Abstract

The application provides a beef cattle abnormal behavior grading recognition method and system, and relates to the technical field of target detection, wherein the method comprises the following steps: acquiring group video scene data of beef cattle, and performing target recognition and fighting behavior recognition on the group video scene data to obtain corresponding target recognition results and preliminary behavior recognition results; when the preliminary behavior recognition results prompt that there is no fighting behavior, dynamic feature extraction is performed in combination with the target recognition results to obtain corresponding dynamic features; when the target recognition results meet preset discrimination rules, static feature extraction is performed in combination with the group video scene data to obtain corresponding static features; and abnormal recognition is performed on the static features and dynamic features respectively to obtain corresponding abnormal recognition results; the application simultaneously combines static sign diagnosis, dynamic gait analysis and time sequence behavior recognition mechanism to realize rapid screening and accurate positioning of abnormal individuals.

Description

Technical Field

[0001] This invention relates to the field of target detection technology, and more specifically to a method and system for hierarchical identification of abnormal behavior in beef cattle. Background Technology

[0002] Currently, as my country's smart animal husbandry industry develops towards large-scale and intensive operations, real-time monitoring of beef cattle health has become a crucial link in ensuring both economic benefits and animal welfare. Abnormal behaviors in beef cattle (such as non-periodic lameness and abnormal fighting) are usually early external manifestations of disease infection, hoof lesions, or group stress. Currently, most farms still heavily rely on manual, timed patrols for monitoring abnormal beef cattle behavior, which suffers from high missed detection rates, strong subjectivity, and large blind spots in nighttime monitoring, making it difficult to achieve 24 / 7, high-frequency, and refined monitoring.

[0003] In recent years, deep learning-based computer vision technology has been gradually introduced into behavior recognition tasks in animal husbandry. For example, Chinese invention patent application (publication number: CN116543462A, invention title: Method for judging the health status of dairy cows based on video skeleton behavior recognition) discloses a scheme for extracting key points of dairy cow skeletons for behavior recognition; another example is Chinese invention patent application (publication number: CN110741962A, invention title: A method for estrus and health monitoring based on dairy cow behavior recognition algorithm) which discloses a cattle behavior discrimination model based on an end-to-end deep learning network. Although these existing technologies have achieved certain recognition results in laboratory environments, they face insurmountable technical bottlenecks in real beef cattle farming engineering scenarios and applications of mobile inspection robots. First, it's difficult to quickly identify abnormal individuals from the complex visual data of a large herd. Real-world large-scale cattle sheds are extremely large, and systems often need to process intricate group visual signals. Without an efficient filtering mechanism, the system can easily get lost in the massive amount of data on healthy cattle, making it difficult to promptly identify and alert on truly abnormal individual cattle.

[0004] Secondly, indiscriminately processing large groups of cattle globally can easily paralyze the computing power of edge devices. In ultra-large-scale ranch inspections, dozens or even hundreds of cattle may appear simultaneously in the frame. The computing power of the edge computing nodes carried by the inspection robot has a physical limit. If high-load static feature recognition and posture estimation algorithms are executed indiscriminately on all cattle in the global frame, it will cause the edge computing chip's computing power to be instantly overloaded and the processing frame rate to drop sharply. Forcing massive amounts of complex data without initial screening into the deep network can even cause the system to crash, completely negating the purpose of industrial-grade real-time monitoring.

[0005] Therefore, how to provide a method for classifying and identifying abnormal behaviors in beef cattle that can solve the above problems is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] In view of this, the present invention provides a method and system for hierarchical identification of abnormal behaviors in beef cattle, constructing a cascaded processing framework consisting of "global scanning—initial anomaly screening—spatial gating—fine diagnosis—state arbitration—filtering and early warning." This framework achieves on-demand release of computing resources by decoupling and parallelizing individual spatial scale gating with macroscopic dynamic initial screening; simultaneously, it combines static physical sign diagnosis, dynamic gait analysis, and temporal behavior recognition mechanisms to achieve rapid screening and accurate localization of abnormal individuals. Furthermore, the present invention deeply couples prior features of animal pathology with a bidirectional temporal neural network and constructs a multi-label state arbitration mechanism and anomaly filtering and evidence storage mechanism based on a state dictionary, thereby achieving low-latency identification and risk classification of individuals exhibiting lameness, fighting, and abnormal static physical signs in complex mobile inspection scenarios.

[0007] To achieve the above objectives, the present invention adopts the following technical solution: A method for hierarchical identification of abnormal behaviors in beef cattle includes the following steps: Acquire group video scene data of beef cattle, and perform target recognition and fighting behavior recognition on the group video scene data to obtain the corresponding target recognition results and preliminary behavior recognition results; The preliminary behavior recognition results are judged, and an alarm is triggered when the preliminary behavior recognition results indicate that fighting behavior exists; When the preliminary behavior recognition result indicates that there is no fighting behavior, dynamic feature extraction is performed in conjunction with the target recognition result to obtain the corresponding dynamic features; When the target recognition result meets the preset discrimination rule, static feature extraction is performed in combination with the group video scene data to obtain the corresponding static features; Anomaly identification is performed on the static and dynamic features respectively to obtain the corresponding anomaly identification results.

[0008] Preferably, the specific processing steps for obtaining the corresponding target recognition results and preliminary behavior recognition results include: Acquire group video scene data of beef cattle, and preprocess the group video scene data to obtain the corresponding current image frame; A preset time detection step size is used to perform target behavior detection on the current image frame to obtain the corresponding preliminary behavior recognition result; Perform cattle target detection on the current image frame and generate corresponding two-dimensional bounding box information; The independent spatiotemporal motion trajectory of each target is extracted based on the two-dimensional bounding box information, and the independent spatiotemporal motion trajectory is used as the target recognition result.

[0009] Preferably, the specific processing steps for obtaining the corresponding static features include: Feature extraction is performed on the current image frame to obtain the corresponding static features.

[0010] Preferably, the specific processing steps for obtaining the corresponding dynamic features include: The independent spatiotemporal motion trajectory is extracted to obtain the corresponding two-dimensional bounding box information; Based on the two-dimensional bounding box information, the macroscopic kinematic physical quantities of cattle under continuous time windows are calculated. The macroscopic kinematic physical quantities are subjected to dynamic analysis and screening to obtain the corresponding target screening results; A lightweight pose estimation model is constructed, and the target selection results are input into the lightweight pose estimation model for processing to obtain the coordinates of key physical nodes of each target. Based on the coordinates of key physical nodes, the local fine-grained gait biomechanical parameters of the target are analyzed, and the local fine-grained gait biomechanical parameters and the coordinates of key physical nodes are used as dynamic features.

[0011] Preferably, the specific processing procedure for anomaly identification of the dynamic features includes: The local fine-grained gait biomechanical parameters and the coordinates of key physical nodes throughout the body are preprocessed to obtain the corresponding multidimensional temporal feature tensors. A discriminant model is constructed, and the multidimensional time-series feature tensor is input into the discriminant model for processing to obtain the corresponding risk confidence. An individual is considered to be in an abnormal state when the risk confidence level meets the threshold requirement.

[0012] Preferably, the discrimination model is a bidirectional long short-term memory network.

[0013] Preferably, the specific processing procedure for anomaly identification of the static features includes: A multi-target detection model is constructed, and the static features are input into the multi-target detection model for processing to obtain the corresponding static feature anomaly discrimination results.

[0014] This invention also provides a graded identification system for abnormal behavior in beef cattle, comprising: The recognition module is used to acquire group video scene data of beef cattle, and to perform target recognition and fighting behavior recognition on the group video scene data to obtain the corresponding target recognition results and preliminary behavior recognition results. The discrimination module is used to discriminate the target recognition result and the preliminary behavior recognition result respectively, and to issue an alarm when the preliminary behavior recognition result indicates that there is fighting behavior; The first feature extraction module is used to perform dynamic feature extraction in conjunction with the target recognition result when the preliminary behavior recognition result indicates that there is no fighting behavior, so as to obtain the corresponding dynamic features. The second feature extraction module is used to construct a multi-target detection model when the target recognition result meets the preset discrimination rules, and to obtain the corresponding static features using the multi-target detection model. The anomaly detection module is used to identify anomalies in the static and dynamic features respectively, and obtain the corresponding anomaly detection results.

[0015] As can be seen from the above technical solution, compared with the prior art, the present invention discloses a method and system for hierarchical identification of abnormal behavior in beef cattle, which has the following beneficial effects: (1) Rapid screening of abnormal individuals in a large-scale cattle herd environment is achieved through a global fighting direct calculation mechanism and an individual-level dynamic coarse screening mechanism; (2) Realize on-demand release of computing power through spatial scale gating mechanism to avoid invalid inference caused by low-quality targets at a distance; (3) Achieve full coverage monitoring of abnormal conditions in beef cattle through a dual-branch heterogeneous diagnostic structure of static physical signs and dynamic gait; (4) Improve the accuracy of limping anomaly identification by deeply fusing pathological prior features with BiLSTM temporal network; (5) A unified risk expression under complex abnormal states is achieved through a multi-label state arbitration mechanism; (6) By using anomaly filtering and automatic evidence storage mechanisms, the rendering overhead of normal targets is greatly reduced, and the long-term operational stability of the edge platform is improved. (7) Improve the efficiency of tracing abnormal individuals and the application value of smart ranch projects by adopting the evidence preservation mechanism of "instant inspection and storage". Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0017] Figure 1 This is an overall flowchart of a method for classifying and identifying abnormal behaviors in beef cattle, provided in an embodiment of the present invention. Figure 2 This is a diagram illustrating the step-by-step screening process from cattle herds to individuals, provided in an embodiment of the present invention. Figure 3 This is a schematic diagram of the structure of the inspection robot provided in an embodiment of the present invention; Figure 4 This is a schematic diagram of the distribution and biomimetic topological mapping of 17 key physical nodes of beef cattle provided in an embodiment of the present invention; Figure 5 A logical sequence diagram for risk confidence threshold discrimination and automatic capture of panoramic evidence provided in the embodiments of the present invention; Figure 6 This is a diagram illustrating the architecture for temporal feature enhancement and abnormal behavior identification in beef cattle provided in this embodiment of the invention. Figure 7 The present invention provides a structural principle block diagram of a graded identification system for abnormal behavior in beef cattle. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] See Figure 1 As shown in the figure, this invention discloses a method for hierarchical identification of abnormal behavior in beef cattle, including the following steps: Acquire group video scene data of beef cattle, and perform target recognition and fighting behavior recognition on the group video scene data to obtain the corresponding target recognition results and preliminary behavior recognition results; The preliminary behavior recognition results are judged, and an alarm is triggered when the preliminary behavior recognition results indicate that fighting behavior exists; When the preliminary behavior recognition result indicates that there is no fighting behavior, dynamic feature extraction is performed in conjunction with the target recognition result to obtain the corresponding dynamic features; When the target recognition result meets the preset discrimination rule, static feature extraction is performed in combination with the group video scene data to obtain the corresponding static features; Anomaly identification is performed on the static and dynamic features respectively to obtain the corresponding anomaly identification results.

[0020] Specifically, in this embodiment of the invention, an inspection robot can be used to collect group video scene data of beef cattle. During use, the inspection robot autonomously patrols within the cattle shed, acquiring continuous video streams. The structure of the inspection robot can be found in [reference needed]. Figure 3 As shown, it includes a mobile chassis module, a visual perception module, a gimbal execution module, a lidar and sensor module, an edge computing and computing power scheduling module, and a wireless communication module; The mobile chassis module drives the inspection robot to move autonomously along a preset inspection route; the visual perception module is installed on the gimbal execution module to collect continuous video streams of beef cattle herds in the pasture; the gimbal execution module is used to adjust the pitch angle and horizontal orientation of the visual perception module to achieve target tracking and field of view optimization; the lidar and sensor module is used to perceive surrounding environmental information, build an environmental map, and assist the robot in autonomous navigation; the edge computing and computing power scheduling module is used to execute the overall video processing flow proposed in this embodiment of the invention; the wireless communication module is used to upload abnormal alarm information, target identity information, and video evidence to the breeding management terminal.

[0021] Reference Figure 2 As shown in the illustration, this embodiment of the invention further demonstrates the step-by-step screening process from herd to individual through visualization. In the initial stage, the input is a video scene containing a large herd of beef cattle, where all targets exist as bounding boxes. Subsequently, a spatial scale gating mechanism performs the first round of screening of the herd targets, retaining only individuals at close range with clear imaging conditions, thereby eliminating distant, low-resolution targets. Based on this, a macroscopic dynamics screening mechanism is further combined to evaluate the target motion characteristics, gradually eliminating individuals that conform to normal motion patterns, significantly reducing the size of the candidate set. Finally, only a very small number of individuals meeting the abnormal motion judgment criteria are retained and marked in red as suspected abnormal cattle individuals.

[0022] In a specific embodiment, the specific processing steps for obtaining the corresponding target recognition result and preliminary behavior recognition result include: Acquire group video scene data of beef cattle, and preprocess the group video scene data to obtain the corresponding current image frame; A preset time detection step size is used to perform target behavior detection on the current image frame to obtain the corresponding preliminary behavior recognition result; Perform cattle target detection on the current image frame and generate corresponding two-dimensional bounding box information; The independent spatiotemporal motion trajectory of each target is extracted based on the two-dimensional bounding box information, and the independent spatiotemporal motion trajectory is used as the target recognition result.

[0023] Specifically, in this embodiment of the invention, the spatial location of the target is predicted using Kalman filtering on the current image frame, and then matched with the intersection-over-union (IoU) ratio between the detected bounding box and the predicted bounding box, thereby assigning a unique digital identity (ID) to each cattle. In complex environments with dense herds and frequent occlusion, when the target is temporarily occluded or lost, the system maintains the continuity of the ID by relying on the predicted trajectory, thereby constructing an independent spatiotemporal motion sequence for each individual on the time axis, realizing a structured decomposition from the group video stream to the individual trajectory stream.

[0024] Furthermore, embodiments of the present invention can also preprocess the current image frame, the specific process of which is as follows: First, a proportional edge padding and clipping process (letterbox processing) is performed to maintain physiological proportions. This involves scaling the target area proportionally based on its longest side, and then filling any insufficient edge portions with solid color pixels for symmetrical centering. The formulas for calculating the minimum scaling factor s and the pixel center symmetry offset (Δx, Δy) are as follows:

[0025] In the formula, T represents the standard input tensor size preset by the neural network; h and ω represent the height and width of the bounding box of the original suspected abnormal individual cropped from the global image, respectively; s is the minimum scaling factor that maintains the proportion of the original physiological features; Δx and Δy represent the center-symmetric offset in the width and height directions of the target canvas, respectively, used to determine the starting position of the image filling.

[0026] In a specific embodiment, the specific processing steps for obtaining the corresponding static features include: Feature extraction is performed on the current image frame to obtain the corresponding static features.

[0027] Specifically, the preset discrimination rule for triggering the extraction of static features is as follows: This invention introduces an adaptive spatial scale gating mechanism to calculate the proportion of the area information of an individual's two-dimensional bounding box in the current image frame in real time. The specific expression is as follows: γ i =Area i / Area total In the formula, Area i This represents the area information of the two-dimensional bounding box. Area total This represents the area of ​​the current image frame.

[0028] When γ i When γ > 0.08, the dual branches of static vital sign diagnosis and dynamic pose calculation are activated. i When the value is ≤0.08, the calculation of distant fuzzy targets is blocked.

[0029] In a specific embodiment, the specific processing steps for obtaining the corresponding dynamic features include: The independent spatiotemporal motion trajectory is extracted to obtain the corresponding two-dimensional bounding box information; Based on the two-dimensional bounding box information, the macroscopic kinematic physical quantities of cattle under continuous time windows are calculated. The macroscopic kinematic physical quantities are subjected to dynamic analysis and screening to obtain the corresponding target screening results; A lightweight pose estimation model is constructed, and the target selection results are input into the lightweight pose estimation model for processing to obtain the coordinates of key physical nodes of the whole body for each target. Based on the coordinates of key physical nodes of the whole body, the local fine-grained gait biomechanical parameters of the target are analyzed, and the local fine-grained gait biomechanical parameters and the coordinates of key physical nodes of the whole body are used as dynamic features. The lightweight pose estimation model can be YOLOv11n-pose.

[0030] Specifically, embodiments of the present invention construct a dual filtering mechanism for rapid removal of normal individuals, which includes: (1) High-speed motion exemption mechanism: When the average moving speed of an individual exceeds the preset running threshold, the system determines that the individual is in a normal brisk walking or running state. The vertical shaking caused by the group's motion characteristics is regarded as a normal dynamic response, thereby directly triggering the exemption and skipping the subsequent calculation.

[0031] (2) Static / Micro-motion Interception Mechanism: Based on the running exemption, the system further calculates the actual physical displacement. When the displacement is lower than the minimum threshold, it is determined that the individual is in a static or local behavior state (such as eating, looking down, etc.), and "displacement interception" is directly executed without entering the high computing power module.

[0032] Ultimately, only individuals meeting the following criteria are retained for the next stage: not exempted, exhibiting valid displacement, and simultaneously possessing either a vertical mutation or a horizontal anomalousness. This mechanism enables the rapid screening of suspected anomalous candidates from the population.

[0033] Based on a dual filtering mechanism for "rapid removal of the normal population," macroscopic dynamic analysis is performed on each tracked individual. This involves extracting the geometric center point sequence of the target bounding box and calculating kinematic characteristics within a preset time window, including average horizontal velocity, true displacement distance, horizontal motion variance, and vertical oscillation amplitude. The geometric center point sequence (x) of the two-dimensional bounding box information is then extracted. k ,y k The corresponding average moving speed V avg With physical displacement D abs The calculation formula is as follows:

[0034] In the formula, △t represents the duration of the observation time window set by the system (i.e., the amount of time change), and represents the effective time span for extracting individual motion sequences for dynamic feature calculation; M represents the total number of frames of the target center point coordinate sequence within the set observation time window △t; (xM ,y M (x1, y1) represents the physical spatial coordinates of the geometric center point of the target's two-dimensional bounding box in the last frame (i.e., the Mth frame) of the observation time window, representing the end position within that time period; (x1, y1) represents the physical spatial coordinates of the geometric center point of the target's two-dimensional bounding box in the first frame (initial frame) of the observation time window, representing the start position within that time period; k represents the current frame index number in the spatiotemporal sequence, used to traverse the instantaneous coordinates (x1, y1) of the target's center point frame by frame. k ,y k ), whose range is the set of constants 1 < k < M.

[0035] If V avg >Threshold run (Threshold) run The abnormal running exemption threshold is defined as the average speed limit for determining whether a cow is walking or running. The exemption mechanism is automatically triggered if the cow is deemed to be running normally. abs Threshold static (Threshold) static The minimum displacement interception threshold (defined as the minimum displacement limit value required to determine whether a cow has actually moved) is used to perform "displacement interception" to remove stationary individuals when the judgment condition is met. This is only true when V is simultaneously satisfied. avg ≤Threshold run D abs ≥Threshold static When there is a sudden vertical drop or abnormal horizontal deviation, it is identified as a suspected target.

[0036] Subsequently, a lightweight pose estimation model (such as YOLOv11n-pose) was applied to the suspected anomalous target to extract the coordinates of 17 key physical nodes, including the head and neck, spine, limbs, and hoof structures, such as... Figure 4 As shown in the figure. Subsequently, spatial translation normalization was performed with the spinal nodes as the origin, and scale normalization was performed with the body axis length to obtain a standardized skeletal structure. Based on this structure, three types of biomechanical parameters were calculated: head and neck pitch rhythm parameters, back arching angle parameters, and longitudinal stride asymmetry parameters. These parameters, together with the node coordinates, constitute a high-dimensional behavioral representation vector for subsequent temporal recognition.

[0037] In a specific embodiment, the specific processing procedure for anomaly identification of the dynamic features includes: The local fine-grained gait biomechanical parameters and the coordinates of key physical nodes throughout the body are preprocessed to obtain the corresponding multidimensional temporal feature tensors. A discriminant model is constructed, and the multidimensional time-series feature tensor is input into the discriminant model for processing to obtain the corresponding risk confidence. An individual is considered to be in an abnormal state when the risk confidence level meets the threshold requirement.

[0038] In one specific embodiment, the discriminant model is a bidirectional long short-term memory network (BiLSTM). BiLSTM captures the time dependence, phase change, and asymmetric mutation features of motion trajectories through bidirectional sequence learning, and finally outputs the confidence level of abnormal behavior through Softmax.

[0039] For details, see Figure 5 As shown, the specific process for determining an individual's abnormal state when the risk confidence level meets the threshold requirement also includes: To meet the engineering application requirements of anomaly detection, this embodiment of the invention adopts a high-sensitivity threshold strategy, which triggers anomaly determination when the confidence level exceeds 0.3, thereby enabling the priority capture of weakly anomalous individuals in the population.

[0040] Meanwhile, to reduce the rendering load of group visualization, this embodiment of the invention constructs a filtering mechanism based on a state dictionary, the core logic of which is as follows:

[0041] Among them, S i (t) This represents the behavioral diagnostic status label (e.g., normal, lame, fighting, or abnormal saliva) of the i-th cow at time t, which is aggregated by the system. i (t) This is a rendering control bit.

[0042] During the visualization phase, the system iterates through the entire group, focusing only on the Triggers. i (t) Abnormal individuals with a value of 1 are highlighted and have their bounding boxes rendered, while the rest of the normal population is automatically filtered out, thus significantly reducing the rendering overhead of the population.

[0043] Simultaneously, it automatically retrieves three consecutive three-second video clips containing the target, generates structured abnormal evidence, and pushes it to the management terminal, realizing a closed-loop processing flow from "group monitoring" to "individual diagnosis".

[0044] For details, see Figure 6 As shown, the specific processing steps to obtain the corresponding multidimensional temporal feature tensor include: To unify video sequences of different lengths, a temporal uniform sampling mechanism is introduced to map variable-length sequences to fixed-length sequences. Let the total number of frames in the current buffer queue be M, and the number of fixed-length frames in the model's standard input be N. The formula for calculating the original video frame index τ(k) corresponding to the k-th standard time step is:

[0045] Time axis alignment is achieved through equidistant sampling, and the individual sequences selected from the population are uniformly standardized.

[0046] Subsequently, the key point coordinates and biomechanical parameters are stitched together frame by frame to form a 37-dimensional temporal feature tensor, which is then input into a BiLSTM network for temporal modeling.

[0047] In a specific embodiment, the specific process for anomaly identification of the static features includes: A multi-target detection model is constructed, and the static features are input into the multi-target detection model for processing to obtain the corresponding static feature anomaly discrimination results.

[0048] See Figure 7 As shown, this embodiment of the invention also provides a system for classifying and identifying abnormal behaviors in beef cattle using any of the above embodiments, comprising: The recognition module is used to acquire group video scene data of beef cattle, and to perform target recognition and fighting behavior recognition on the group video scene data to obtain the corresponding target recognition results and preliminary behavior recognition results. The discrimination module is used to discriminate the target recognition result and the preliminary behavior recognition result respectively, and to issue an alarm when the preliminary behavior recognition result indicates that there is fighting behavior; The first feature extraction module is used to perform dynamic feature extraction in conjunction with the target recognition result when the preliminary behavior recognition result indicates that there is no fighting behavior, so as to obtain the corresponding dynamic features. The second feature extraction module is used to construct a multi-target detection model when the target recognition result meets the preset discrimination rules, and to obtain the corresponding static features using the multi-target detection model. The anomaly detection module is used to identify anomalies in the static and dynamic features respectively, and obtain the corresponding anomaly detection results.

[0049] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to the method section.

[0050] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for hierarchical identification of abnormal behaviors in beef cattle, characterized in that, Includes the following steps: Acquire group video scene data of beef cattle, and perform target recognition and fighting behavior recognition on the group video scene data to obtain the corresponding target recognition results and preliminary behavior recognition results; The preliminary behavior recognition results are judged, and an alarm is triggered when the preliminary behavior recognition results indicate that fighting behavior exists; When the preliminary behavior recognition result indicates that there is no fighting behavior, dynamic feature extraction is performed in conjunction with the target recognition result to obtain the corresponding dynamic features; When the target recognition result meets the preset discrimination rule, static feature extraction is performed in combination with the group video scene data to obtain the corresponding static features; Anomaly identification is performed on the static and dynamic features respectively to obtain the corresponding anomaly identification results.

2. The method for hierarchical identification of abnormal behavior in beef cattle according to claim 1, characterized in that, The specific processing steps to obtain the corresponding target recognition results and preliminary behavior recognition results include: Acquire group video scene data of beef cattle, and preprocess the group video scene data to obtain the corresponding current image frame; A preset time detection step size is used to perform target behavior detection on the current image frame to obtain the corresponding preliminary behavior recognition result; Perform cattle target detection on the current image frame and generate corresponding two-dimensional bounding box information; The independent spatiotemporal motion trajectory of each target is extracted based on the two-dimensional bounding box information, and the independent spatiotemporal motion trajectory is used as the target recognition result.

3. The method for hierarchical identification of abnormal behavior in beef cattle according to claim 2, characterized in that, The specific processing steps to obtain the corresponding static features include: Feature extraction is performed on the current image frame to obtain the corresponding static features.

4. The method for hierarchical identification of abnormal behavior in beef cattle according to claim 2, characterized in that, The specific processing steps to obtain the corresponding dynamic features include: The independent spatiotemporal motion trajectory is extracted to obtain the corresponding two-dimensional bounding box information; Based on the two-dimensional bounding box information, the macroscopic kinematic physical quantities of cattle under continuous time windows are calculated. The macroscopic kinematic physical quantities are subjected to dynamic analysis and screening to obtain the corresponding target screening results; A lightweight pose estimation model is constructed, and the target selection results are input into the lightweight pose estimation model for processing to obtain the coordinates of key physical nodes of each target. Based on the coordinates of key physical nodes, the local fine-grained gait biomechanical parameters of the target are analyzed, and the local fine-grained gait biomechanical parameters and the coordinates of key physical nodes are used as dynamic features.

5. The method for hierarchical identification of abnormal behavior in beef cattle according to claim 4, characterized in that, The specific processing steps for anomaly identification of the dynamic features include: The local fine-grained gait biomechanical parameters and the coordinates of key physical nodes throughout the body are preprocessed to obtain the corresponding multidimensional temporal feature tensors. A discriminant model is constructed, and the multidimensional time-series feature tensor is input into the discriminant model for processing to obtain the corresponding risk confidence. An individual is considered to be in an abnormal state when the risk confidence level meets the threshold requirement.

6. The method for hierarchical identification of abnormal behavior in beef cattle according to claim 5, characterized in that, The discriminant model is a bidirectional long short-term memory network.

7. The method for hierarchical identification of abnormal behavior in beef cattle according to claim 3, characterized in that, The specific processing steps for anomaly identification of the static features include: A multi-target detection model is constructed, and the static features are input into the multi-target detection model for processing to obtain the corresponding static feature anomaly discrimination results.

8. A system utilizing the method for hierarchical identification of abnormal behavior in beef cattle according to any one of claims 1-7, characterized in that, include: The recognition module is used to acquire group video scene data of beef cattle, and to perform target recognition and fighting behavior recognition on the group video scene data to obtain the corresponding target recognition results and preliminary behavior recognition results. The discrimination module is used to discriminate the target recognition result and the preliminary behavior recognition result respectively, and to issue an alarm when the preliminary behavior recognition result indicates that there is fighting behavior; The first feature extraction module is used to perform dynamic feature extraction in conjunction with the target recognition result when the preliminary behavior recognition result indicates that there is no fighting behavior, so as to obtain the corresponding dynamic features. The second feature extraction module is used to construct a multi-target detection model when the target recognition result meets the preset discrimination rules, and to obtain the corresponding static features using the multi-target detection model. The anomaly detection module is used to identify anomalies in the static and dynamic features respectively, and obtain the corresponding anomaly detection results.

Images