Method for analyzing a group of images of the behavior of the state of health of a flock of poultry in a poultry house and system therefor

Abstract

This invention discloses a method and system for analyzing the health status behavior of poultry flocks from images, relating to the fields of computer vision and smart farming technology. The method includes: firstly, acquiring monitoring images and extracting the poultry's location coordinates; secondly, dividing the poultry into an active first target set and an inactive second target set based on displacement features; thirdly, constructing a discrete topological network using the first target set to generate candidate spatial gap regions and identifying edge targets; fourthly, generating a biological rejection intensity index based on the head orientation vector of the edge targets, and determining stable rejection regions based on temporal stability; and fifthly, outputting a health anomaly warning when the center of the stable rejection region and the coordinates of the second target set satisfy spatial matching conditions. This invention effectively solves the problem of distinguishing between physiological quiescence and pathological stagnation in high-density farming by analyzing the group's biological rejection behavior to pinpoint abnormal individuals, thus improving monitoring accuracy.

Description

Technical Field

[0001] This invention relates to the fields of computer vision and smart farming technology, specifically to a method and system for analyzing the health status and behavior of chicken flocks in poultry houses. Background Technology

[0002] With the development of large-scale intensive farming technology, the use of machine vision technology for non-contact monitoring of poultry has become a core trend in smart farming. It is widely used in livestock and poultry growth monitoring, quantity statistics and environmental control, aiming to reduce farming costs and improve management efficiency by replacing manual inspections with automation.

[0003] In the monitoring of poultry health status, existing technologies typically use target detection algorithms to locate individuals and combine optical flow or skeleton analysis techniques to track their movement trajectories and postures. The health status is then assessed by statistically analyzing the individual's movement amplitude, activity frequency, or specific abnormal postures (such as falls or lameness).

[0004] However, in scenarios with high stocking density and variable ambient lighting, the physical characteristics of the data are often insufficient to characterize the physiological state of the organisms. Specifically, healthy poultry will also exhibit prolonged periods of stillness or lying down when resting, sleeping, or digesting. This physiological stillness is almost indistinguishable from pathological stagnation caused by disease in terms of pixel features and motion parameters. If the system relies too much on positive features such as the texture, contour, or displacement of the target itself, it is very easy for the system to misjudge diseased individuals as normal resting individuals if they are in the early latent stage without obvious appearance distortion or simply exhibit slow movement, or to completely miss detection due to occlusion.

[0005] Therefore, how to overcome the limitations of single-target detection in complex scenarios where individual visual features fail or semantic ambiguity exists, and how to discover and utilize the implicit interaction logic between group distribution and environmental space to identify hidden health abnormalities, is a technical challenge that urgently needs to be solved in the field of intelligent poultry monitoring. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides a method and system for analyzing the health status and behavior of chicken flocks in poultry houses using images.

[0007] To achieve the above objectives, the technical solution of the present invention is as follows:

[0008] In a first aspect, this invention discloses a method for group analysis of the health status and behavior of chicken flocks in poultry houses, comprising the following steps:

[0009] Acquire monitoring image data of poultry houses, extract features from the monitoring image data, and obtain the location coordinates of multiple poultry targets;

[0010] The displacement characteristics of each poultry target are calculated based on the preset time window and positioning coordinates, and the poultry targets are divided into a first target set in an active state and a second target set in an inactive state based on the displacement characteristics.

[0011] A discrete topology network is constructed based on the positioning coordinates of the first target set. Mesh cells in the discrete topology network that meet the preset spatial sparsity conditions are aggregated to generate candidate spatial gap regions. Edge targets in the first target set located on the topological boundary of the candidate spatial gap regions are determined.

[0012] Based on the head orientation vector of the edge target extracted from the surveillance image data, and the biological repulsion strength index of the candidate spatial void region is generated according to the radial geometric relationship between the head orientation vector and the center point of the candidate spatial void region.

[0013] Candidate spatial void regions whose biological repulsion strength index meets the preset temporal stability condition are identified as stable repulsion regions.

[0014] In response to the fact that the center point of the stable exclusion region and the positioning coordinates of the second target set meet the preset spatial matching conditions, a health abnormality warning signal for the corresponding poultry target is output.

[0015] Secondly, this invention discloses a poultry house chicken flock health status behavior image group analysis system, including:

[0016] The image acquisition module is used to acquire monitoring image data of poultry houses and extract features from the monitoring image data to obtain the location coordinates of multiple poultry targets.

[0017] The state analysis module is used to calculate the displacement characteristics of each poultry target based on a preset time window and positioning coordinates, and to divide the poultry targets into a first target set in an active state and a second target set in an inactive state based on the displacement characteristics.

[0018] The topology analysis module is used to construct a discrete topology network based on the positioning coordinates of the first target set, aggregate the grid cells in the discrete topology network that meet the preset spatial sparsity conditions, generate candidate spatial gap regions, and determine the edge targets in the first target set located on the topological boundary of the candidate spatial gap regions.

[0019] The repulsion strength calculation module is used to extract the head orientation vector of the edge target based on the monitoring image data, and generate the bio-repulsion strength index of the candidate spatial gap region according to the radial geometric relationship between the head orientation vector and the center point of the candidate spatial gap region.

[0020] The stability determination module is used to identify candidate spatial void regions whose biological repulsion strength index meets the preset temporal stability conditions as stable repulsion regions.

[0021] The abnormality warning module is used to output a health abnormality warning signal for the corresponding poultry target when the center point of the stable exclusion area and the positioning coordinates of the second target set meet the preset spatial matching conditions.

[0022] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0023] 1. By constructing a discrete topological network and quantifying the biological repulsion intensity index of marginal individuals, the system can accurately capture the instinctive avoidance behavior of surrounding healthy groups towards disease targets. This group-to-individual analytical dimension enables the system to accurately pinpoint the true source of disease that triggers group panic or avoidance, even in the absence of obvious individual appearance distortions, by utilizing the mutually exclusive characteristics of group behavior. This significantly improves the biological interpretability and accuracy of monitoring results.

[0024] 2. After generating candidate gap regions, a morphological topology diversion and linear parallel flow verification mechanism is introduced. By calculating the aspect ratio of the gap and the motion velocity vector of the edge individuals and the parallel flow interference index of the gap's long axis, the system can intelligently identify and eliminate false exclusion regions formed by facility obstruction or group feeding. This enables it to be widely adapted to aquaculture scenarios with different structures and levels of automation, eliminating the need for frequent manual setting of shielding areas and greatly reducing the interference of environmental noise on the monitoring system.

[0025] 3. A reverse localization strategy is adopted, which does not directly search for occluded diseased individuals, but detects significant topological voids formed by healthy groups. Since the rejection behavior of healthy groups usually forms physical spaces larger than the individual scale, this void feature is more robust and visible than individual features. As long as the group has avoidance behavior, the system can lock the stable rejection area and, combined with the adaptive spatial matching radius, accurately capture cases located in the center of the void or even partially occluded, effectively preventing the missed detection of core pathogens.

[0026] 4. In the behavioral dimension, the dot product analysis of the head orientation vector and the radial centrifugal vector confirms that the gaps are formed by repulsion rather than random dispersion. In the temporal dimension, the mean and variance of the repulsion index are statistically analyzed using a sliding window to filter out instantaneous gaps caused by sudden fright. In the spatial dimension, it is required that the repulsion center must physically collide with an inactive individual to trigger an alarm. This multimodal logical closed loop ensures that every warning has sufficient conditions, providing highly reliable decision support for unmanned precision farming. Attached Figure Description

[0027] 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 some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0028] Figure 1 This is an overall block diagram of the method in Embodiment 1 of the present invention;

[0029] Figure 2 This is a flowchart illustrating the overall execution process of the method in Embodiment 1 of the present invention.

[0030] Figure 3 This is an overall block diagram of the system in Embodiment 2 of the present invention. Detailed Implementation

[0031] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. 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.

[0032] Application Overview: In the field of modern intensive poultry farming, especially in the health monitoring of high-density free-range chicken flocks, the spatial representation of group biological interaction behavior is regarded as a deep-seated key indicator for measuring the biosafety of the farming environment and the health status of individuals. The accurate determination of such abnormal health events is essentially a dynamic generation process of biological repulsion force field at the animal behavior level. That is, when an individual exhibits pathological abnormalities (such as death or even highly infectious symptoms), healthy neighboring individuals will be driven by instinctive fear or hygiene avoidance mechanisms to generate radial displacement away from the source of the abnormality, thereby forcibly tearing out a spatial gap with specific vector characteristics in the originally uniformly distributed high-density group topology.

[0033] However, existing technologies lack a mechanism to verify the causal logic between an individual's static state and the rejection response of the group, making it impossible to accurately identify physiological stillness confusion and environmental noise interference in complex farming environments. Physiological stillness is manifested in poultry remaining motionless for extended periods due to deep sleep or resting on the ground, but the surrounding group does not reject them and may even cling to them closely. Environmental noise interference, on the other hand, manifests as physical gaps caused by automated feeding chains or uneven lighting. As a result, a strict semantic correspondence cannot be established between the movement characteristics of the target individual and the topological distribution of the surrounding group, leading to misjudgments or logical breaks in the monitoring system's assessment of static target attributes, which in turn affects the accuracy of qualitative identification of the source of core diseases and the timeliness of early warning.

[0034] If the above problems are not addressed, the health monitoring system will continuously lose its ability to objectively identify high-risk abnormal targets. Specifically, the failure to effectively eliminate physiological quiescence will lead to alarm fatigue, making it difficult for the management system to focus on the real risk points, thus weakening the signal-to-noise ratio of health warnings. Simultaneously, the failure to correctly attribute pathological gaps caused by herd rejection will result in the long-term retention of core pathogens, allowing diseases to spread covertly within the farm and ultimately causing the collapse of the herd's health defenses. Therefore, inaccurate monitoring feedback will systematically hinder farms from achieving refined biosecurity management, affecting the final survival rate and economic benefits of livestock.

[0035] Example 1:

[0036] like Figures 1-2 As shown, the method for group analysis of the health status behavior of chickens in a poultry house includes the following steps:

[0037] Step S1: Acquire monitoring image data of the poultry house, and extract features from the monitoring image data to obtain the location coordinates of multiple poultry targets;

[0038] Specifically, the system first acquires real-time monitoring image data through industrial-grade high-definition cameras deployed on the top of the poultry house. This image data is typically high-resolution overhead video frames to minimize occlusion between individuals.

[0039] After acquiring surveillance image data, the system extracts features to obtain location coordinates. This process integrates deep learning object detection and keypoint regression techniques at the underlying logic level. Specifically, this embodiment pre-builds and trains a multi-task convolutional neural network model, which uses a lightweight backbone network (such as MobileNetV3 or EfficientNet) to adapt to the real-time requirements of edge computing devices. The model's input layer receives normalized surveillance image frames, and the intermediate layer fuses multi-scale features through a Feature Pyramid Network (FPN) to simultaneously capture large-sized adult chickens and small-sized chicks or partially occluded targets. The model's output head is designed with a dual-branch structure: the classification regression branch outputs the bounding box and confidence score for each poultry target, while the keypoint regression branch outputs the head keypoints and body center keypoints of the poultry.

[0040] In the actual inference process, the system inputs the current frame image into the model, and the model first outputs the bounding box information of all detected poultry targets. To eliminate redundant detection boxes, the system uses the non-maximum suppression (NMS) algorithm to filter out the optimal bounding boxes based on a preset intersection-over-union (IoU) threshold. Subsequently, the system extracts the localization coordinates of the poultry targets based on the detected bounding boxes. In this embodiment, the localization coordinates are not simply taken as the geometric center of the bounding box, but rather combined with body keypoint information. Specifically, the system uses the coordinates of the body center keypoint output by the keypoint regression branch as the representative position of the poultry target in physical space, denoted as . Simultaneously, the system also extracts and stores the coordinates of key points in the head. These two key points are not only used for subsequent spatial positioning, but also provide the necessary data foundation for calculations in subsequent steps.

[0041] To further enhance the robustness of localization, especially addressing potential issues like equipment obstruction or uneven lighting within poultry houses, this embodiment also introduces a temporal correlation mechanism. Although step S1 primarily focuses on single-frame detection, in actual engineering implementation, the system maintains a short-time memory queue. For detection boxes in the current frame with low confidence but spatially highly overlapping with high-confidence targets in previous frames, the system utilizes a Kalman filter algorithm to smooth their coordinates, ensuring that the output localization coordinate sequence is smooth and continuous.

[0042] Finally, after the above processing, the system outputs a localization dataset of all valid poultry targets in the current frame. This dataset includes, but is not limited to, a unique identifier (ID) and precise body center coordinates for each target. and the coordinates of key points in the head These high-precision coordinate data form the basis for subsequent calculations of displacement features and construction of topological networks, thus realizing a complete transformation from raw pixel data to structured spatial data and laying a solid data foundation for the entire group behavior analysis.

[0043] Step S2: Calculate the displacement characteristics of each poultry target based on the preset time window and positioning coordinates, and divide the poultry targets into a first target set in an active state and a second target set in an inactive state based on the displacement characteristics;

[0044] After the poultry target is accurately located in step S1, the purpose of step S2 is to perform state sorting of the group based on time-series motion data. This step acts as an initial screening filter, aiming to remove individuals that have been quiescent for a long time due to pathological or physiological reasons from the nodes that construct the topological network, thereby creating the topological gaps required for subsequent steps.

[0045] In the specific implementation of this embodiment, the system first sets a preset time window for state evaluation. (For example, The time window is set to be between 10 and 30 seconds. The selection of this time window needs to balance the real-time performance of the calculations with the stability of the state judgment. A window that is too short may be susceptible to transient noise interference, while a window that is too long will cause the system to lag in responding to sudden anomalies. Based on this time window, the system extracts the location coordinates of each poultry target in consecutive frames of images from the historical data cache, constructing a location coordinate sequence. Assuming the current time is t and the sampling frame rate is f, then for any poultry target i, its location coordinate sequence can be expressed as: ,in The total number of frames within the window. This represents the coordinates of the body center of target i in the kth frame.

[0046] Subsequently, the system calculates the displacement characteristics of the target based on the positioning coordinate sequence. To accurately reflect the individual's true activity level over a period of time (including small-scale pacing, feeding swaying, etc.), this embodiment does not simply calculate the straight-line distance between the start and end frames of the window, but rather calculates the sum of the Euclidean distances between adjacent frames in the sequence, defining this as the cumulative displacement of the poultry target. The specific calculation formula is as follows:

[0047] ;

[0048] in, Let represent the cumulative displacement of the i-th poultry target. This refers to the Euclidean distance between two adjacent frames. This calculation method can effectively capture the characteristics of healthy individuals whose overall position has not changed significantly, but whose limbs or bodies maintain continuous micro-movements.

[0049] After obtaining the cumulative displacement of all targets, the system introduces a preset active determination threshold. Perform binary classification. This threshold is typically set dynamically based on the statistical distribution of historical normal active data (e.g., taking the lower quartiles). The specific classification logic is as follows: If... This indicates that the target maintained a high level of movement within the time window, and the system classifies it as an active individual and includes it in the first target set (denoted as...). );like This indicates that the target was essentially in a static or minimally moving state within the time window (possibly resting, in deep sleep, or pathologically paralyzed). The system classifies it as an inactive individual and categorizes it into the second target set (denoted as...). ).

[0050] For example, assume a preset activity determination threshold. The value is 50 pixels. If target A's cumulative displacement within 30 seconds is 300 pixels (frequent movement), it is assigned to the first target set; if target B's cumulative displacement within the same time period is only 15 pixels (only slight center of gravity shift), it is assigned to the second target set. Through this hard partitioning based on temporal displacement characteristics, the system successfully separates the group into the foundation for building the network (the first target set) and potential excluded objects (the second target set), providing data input for subsequent construction of discrete topology networks and generation of anomalous gaps.

[0051] Step S3: Construct a discrete topology network based on the positioning coordinates in the first target set, aggregate the grid cells in the discrete topology network that meet the preset spatial sparsity conditions, generate candidate spatial gap regions, and determine the edge targets in the first target set located on the topological boundary of the candidate spatial gap regions;

[0052] After filtering out the first set of active targets in step S2 using displacement features, step S3 aims to utilize the spatial distribution characteristics of these active individuals to reverse-engineer anomalies in the population by constructing a mathematical geometric model. This embodiment does not employ the traditional heatmap method based on pixel density; instead, it constructs a discrete topological network based on the location coordinates of the first target set and uses computational geometry methods to accurately define the candidate spatial gap regions.

[0053] Specifically, the system first extracts the set of location coordinates of all poultry targets in the first target set. ,in Let be the total number of active individuals. Using the Delaunay triangulation algorithm, connect the above coordinates to generate a triangular mesh covering the entire poultry flock. This algorithm ensures that the generated triangular mesh cells have the property of maximizing minimum angles, thus avoiding the formation of elongated triangles and guaranteeing that the mesh structure accurately reflects the neighborhood topology between individuals. The generated triangular mesh is denoted as . ,in This represents the k-th triangular mesh cell, consisting of the coordinate vertices of three active individuals. constitute.

[0054] Subsequently, the system performs a quantitative analysis of the spatial sparsity of the mesh. First, it calculates the geometric area of ​​all triangular mesh cells in the triangular mesh diagram. For any given triangular mesh cell… Its area The following can be calculated using vertex coordinates:

[0055] ;

[0056] After obtaining the area set of all grid cells Then, the system calculates the median of the set. To accommodate the differences in chicken size at different growth stages and variations in camera installation height, the system introduces adaptive gap determination logic, multiplying the median by a preset magnification factor. (For example, The value ranges from 2.5 to 4.0, resulting in a dynamic void area threshold. .

[0057] coefficient here The value of has clear biostatistical significance. Under natural conditions, the social distance of individual poultry follows a specific probability distribution. When the local grid area exceeds 2.5 times the median, it significantly exceeds the physiological space required for normal heat dissipation or movement, likely indicating the presence of a coercive biological repulsion force. Setting it too low can lead to false detection of loose gaps caused by random movement of individuals; if If the setting is too high, tiny gaps in the early stages of rejection may be missed.

[0058] Based on this threshold, the system performs filtering and aggregation operations. First, it iterates through all grid cells and aggregates the areas. The triangles are labeled as sparse grid cells, that is, grid cells that meet the preset spatial sparsity conditions. These sparse grid cells physically correspond to local open areas formed in the group due to some reason (such as avoidance). Then, using the Connected Component Labeling algorithm, sparse grid cells that are spatially adjacent (i.e., share at least one common edge) are merged, and each independent connected domain constitutes a candidate spatial gap region.

[0059] In this process, to avoid misclassifying the background area located between the outer contour of the entire poultry flock and the environmental boundary as internal voids, the system also needs to perform global boundary culling. Specifically, the system identifies triangles located at the edge of the convex hull in the triangular mesh diagram. If any side of the triangle exceeds the preset physical connectivity limit (e.g., 5 times the average body length of poultry), it is directly identified as the environmental background and not marked as a sparse mesh cell, thereby ensuring that the generated candidate spatial void regions are limited to enclosed voids within the flock.

[0060] Finally, the system needs to determine the key reference objects used for subsequent repulsion force calculations, namely, the edge targets. For each generated candidate spatial void region, the system identifies its topological boundary. Specifically, the system retrieves the edges of all triangular mesh cells constituting the region. If an edge belongs to only one triangle within the region (i.e., the other side of the edge is a non-sparse region or has no mesh), then the edge is defined as a boundary edge. The poultry targets corresponding to the vertices of all boundary edges are identified as edge targets located on the topological boundary of the candidate spatial void region in the first target set. The system outputs a data structure containing the geometric centers of the candidate spatial void regions and the corresponding edge target index list, which is then transmitted to the next processing stage.

[0061] After generating candidate spatial gap regions in step S3, this embodiment does not directly calculate the repulsion strength for all regions. Instead, it first introduces a morphology-based inspection mechanism. This is because in modern poultry houses, automated feeding chains or water lines often naturally divide the flock into narrow strip-shaped gaps. This physical division is not a biological repulsion behavior, and if it is not distinguished, it can easily lead to false alarms.

[0062] Specifically, the system first calculates the minimum bounding rectangle (MBR) for each candidate spatial gap region generated in step S3. The minimum bounding rectangle is the rectangle that can contain the region and has the smallest area. From this, the system extracts two key geometric parameters: the length of the major axis. With minor axis length The system then calculates the ratio of the two to obtain the aspect ratio characteristic value. :

[0063] ;

[0064] Based on this feature value, the system executes morphological topology splitting logic. The system sets a preset morphological threshold. (For example, set it to 2.5).

[0065] like This indicates that the void is geometrically close to a circle or an irregular block shape, consistent with the characteristics of natural cavities formed by biological repulsion (such as avoiding diseased or dead individuals). The system marks such areas as point-source repulsion areas and directly proceeds to step S4 for this area to perform subsequent calculations of the biological repulsion intensity index.

[0066] like This indicates that the gap exhibits a significant elongated strip-like characteristic, which is highly likely to be linear interference caused by the facility. The system marks such areas as suspected linear interference areas. These areas cannot be simply eliminated directly (because there may be multiple sick chickens dying in clusters), so the system initiates a linear parallel flow verification step for secondary confirmation.

[0067] In the linear parallel flow verification, the system first extracts the major axis direction vector of the minimum bounding rectangle corresponding to the suspected linear interference region. Simultaneously, using the preset time window and positioning coordinates cached in steps S1 and S2, the system calculates the instantaneous velocity vectors of all edge targets in the area at the current moment. For the i-th edge target, its velocity vector... Can be used through current frame coordinates Coordinates of the preceding frame The difference is obtained, where Short frame intervals (e.g., 3-5 frames) are used to calculate instantaneous velocity, in order to filter out single-frame jitter and characterize instantaneous motion trends.

[0068] Subsequently, the system calculates the parallelism between the motion trend of each edge target and the long axis of the gap. Specifically, this is done by calculating the motion velocity vector. with major axis direction vector The parallel flow interference index is obtained by taking the absolute value of the cosine of the angle between the two edges. Next, the system calculates the arithmetic mean of this absolute value for all edge targets to obtain the parallel flow interference index. :

[0069] ;

[0070] in, This represents the total number of edge targets.

[0071] The index The physical meaning is: if most of the chickens at the edge move parallel to the long axis of the gap (i.e., A larger value (close to 1) is typically a characteristic behavior of chickens feeding or moving along the feeding chain, and is considered a non-pathological environmental disturbance. Conversely, if the flock's movement direction is chaotic or perpendicular to its long axis (i.e.,...), it indicates a more serious environmental disturbance. If the size is small, it may still be an effective rejection behavior.

[0072] Based on this, the system will calculate the... Compared with the preset parallel flow determination threshold (For example, 0.8) for comparison:

[0073] like The system determined that the suspected linear interference area was a non-biological rejection area (i.e., confirmed as facility interference). In order to block subsequent processes, the system directly assigned a biological rejection intensity index of zero to the area. If the preset suppression value is used, step S4 will not be calculated again.

[0074] like The system determines that the suspected linear interference region is an effective rejection region (although it is long and narrow, its behavior does not conform to the pipeline characteristics). The system continues to execute step S4 for this region to generate the biological rejection intensity index of the candidate spatial gap region.

[0075] Through this preliminary diversion and verification step, the system effectively filters out common facility interference noise in high-density aquaculture environments, ensuring the effective allocation of computing power for subsequent repulsion intensity calculations and the accuracy of early warning results.

[0076] Step S4: Extract the head orientation vector of the edge target based on the monitoring image data, and generate the biorepulsion intensity index of the candidate spatial gap region according to the radial geometric relationship between the head orientation vector and the center point of the candidate spatial gap region;

[0077] After successfully defining the candidate spatial gap region in physical space and locking its surrounding edge targets in step S3, the purpose of step S4 is to delve deeper into the microscopic behavioral semantics of these edge targets. By quantifying the posture tendency of the population relative to the gap, a biological repulsion strength index characterizing the properties of the gap is generated. This step represents a leap from simple geometric topological analysis to biological behavioral dynamics analysis and is a key step in distinguishing between random dispersion and pathological repulsion.

[0078] In this embodiment, the system first traverses all edge targets identified in step S3. For any edge target i, the system needs to extract its head orientation vector based on the monitoring image data. Although key point information of the entire image may have been initially obtained in step S1, in this step, to obtain higher accuracy, the system adopts an on-demand refinement strategy. Specifically, the system extracts the head orientation vector based on the positioning coordinates of the edge target. A region of interest (ROI) centered at these coordinates is extracted from the original surveillance image data. This ROI image is then input into a pre-trained keypoint regression subnetwork (which focuses on local texture details), outputting the head keypoint coordinates of the edge target in the ROI coordinate system. and coordinates of the body's center key point The keypoint regression subnetwork used in this step is a pre-built lightweight convolutional neural network (e.g., using ResNet-18 as the backbone), with its input layer size configured as follows: The pixels are adapted to the ROI size, and the output layer is a fully connected layer that directly regresses the normalized coordinate offset between the head and the center point. This sub-network is fine-tuned using high-resolution local textures to correct keypoint drift that may occur in step S1 during large field-of-view detection.

[0079] Based on the coordinates of the aforementioned key points, the system constructs a difference vector through vector subtraction, i.e. To eliminate the influence of individual body size (i.e., pixel distance) on orientation determination, the system normalizes the difference vector to obtain a unit head orientation vector. : ;

[0080] in This represents the L2 norm of a vector. It represents the current line of sight or direction of travel of the edge target i.

[0081] It is worth noting that in the actual calculation process, in order to avoid the overlap of key points (i.e. The computational anomaly (approaching zero) was addressed by introducing a small perturbation into the system. (For example The corrected normalization formula is: If the magnitude is lower than the preset validity threshold, the system will automatically discard the invalid vector and not include it in the subsequent repulsion index calculation.

[0082] Next, the system calculates the radial geometric relationship between the head orientation vector and the center point of the candidate spatial gap region. First, the system obtains the geometric center coordinates of the candidate spatial gap region. (These coordinates can be obtained from the arithmetic mean of the vertices of all the grid cells constituting the region). Subsequently, the coordinates from the geometric center are constructed. Point to edge target location coordinates radial centrifugal vector Similarly, normalizing this vector yields the unit radial centrifugal vector. .

[0083] To quantify the degree of repulsion or attraction of an individual relative to the center of the gap, the system calculates the head orientation vector. With unit radial centrifugal vector The dot product is calculated and defined as the centrifugal repulsion factor. :

[0084] ;

[0085] in Let be the angle between the two vectors. When the angle between the head of the edge target and the centrifugal direction is less than 90 degrees, it indicates that the individual is exhibiting a posture of turning its back to the center of the gap and fleeing or avoiding it; when When an individual is facing the center of a gap, it indicates that they may be observing or eating.

[0086] Finally, the system performs weighted aggregation of the centrifugal repulsion factors of all edge targets based on unilateral inhibition logic to generate the final biological repulsion strength index. This logic aims to filter out irrelevant individuals that are centered on the center (i.e., attraction behavior should not offset repulsion behavior), and only accumulates the components that exhibit repulsion characteristics. The specific calculation formula is as follows:

[0087] ;

[0088] in The function represents the total number of edge targets. It functions similarly to the ReLU activation function, setting all non-positive repulsion factors to zero.

[0089] Based on the above calculations, the system outputs the biorepulsion strength index of the candidate spatial void region. The index is a dimensionless value between 0 and 1. The closer the value is to 1, the stronger the centrifugal escape tendency consistently exhibited by the edges of the population surrounding the gap, which is highly likely a biological rejection response to some negative stimulus (such as the smell or unusual appearance of a dead chicken) inside the gap. Conversely, if the value is low, it is more likely that the gap is a randomly formed physical gap. This index provides a quantitative basis for subsequent determination of temporal stability.

[0090] Step S5: The candidate spatial void regions whose biological repulsion strength index meets the preset temporal stability conditions are determined as stable repulsion regions.

[0091] After generating the biological rejection intensity index of the candidate spatial gap region in the single frame image in step S4, step S5 introduces the temporal dimension. By analyzing the persistence and stability of rejection behavior, transient interference caused by fright, fighting or random movement is filtered out, thereby accurately defining the candidate spatial gap region as a stable rejection region.

[0092] In the specific implementation of this embodiment, the system first needs to solve the cross-frame target association problem, that is, to determine whether the candidate gap detected in the current frame and the gap in the previous frame belong to the same observation object. To this end, the system maintains a dynamic list of observation objects, which records the coordinates of the center points of the candidate spatial gap regions identified in the previous frames and their historical repulsion indices. When processing the current frame (let's say frame t), the system obtains the center point of a candidate spatial gap region in the current frame. Iterate through the center points of all candidate gaps recorded in the preceding frame (frame t-1). .

[0093] System Calculation With each of the preceding frames Euclidean distance between To prevent erroneous associations (such as forcibly associating two distant gaps), the system introduces a preset tracking radius. This serves as a threshold for judgment. The system filters out those that meet the criteria. The system identifies all preceding objects and selects the one with the smallest Euclidean distance, establishing it as the same observation object for the current candidate gap. Through this frame-by-frame recursive association mechanism, the system can construct a time series of the biorepulsion intensity index of this observation object in multiple consecutive monitoring images, denoted as […]. , where N is the length of the sliding window used for stability analysis (e.g., 30 frames).

[0094] Subsequently, the system processed the time series. Perform statistical characteristic analysis and calculate its arithmetic mean. and variance The specific calculation formula is as follows:

[0095] ;

[0096] ;

[0097] These two statistics represent the baseline strength and fluctuation level of the rejection behavior, respectively. Based on this, the system introduces a dual-judgment logic:

[0098] Compare arithmetic means Compared with the preset strength benchmark value (For example, set it to 0.6). If This indicates that within the observation window, the group as a whole maintained a high level of rejection.

[0099] Comparing variances Compared with the preset fluctuation benchmark value (For example, set it to 0.05). If This indicates that the repulsion intensity fluctuates very little, ruling out instantaneous high repulsion caused by fright (usually accompanied by rapid decay of high variance) or random dispersion (usually low intensity and large fluctuation).

[0100] Ultimately, only when both of the above conditions are met, i.e. Only then does the system determine that the candidate spatial gap region is not random noise, but a stable repulsion region with biological significance. The geometric center coordinates of this region are locked and passed to subsequent steps for final disease attribution matching.

[0101] Step S6: In response to the fact that the center point of the stable rejection region and the positioning coordinates of the second target set meet the preset spatial matching conditions, output the corresponding poultry target health abnormality warning signal;

[0102] After confirming the existence of a biologically significant stable rejection region in the monitoring scene in step S5, step S6 aims to logically associate this group rejection phenomenon with specific inactive individuals through spatial collision matching, thereby accurately locating the source of the disease and outputting a health abnormality warning signal.

[0103] In the specific implementation of this embodiment, the system first needs to define a metric for determining spatial overlap, namely the matching determination radius. To enable this radius to adapt to changes in body size of poultry at different growth ages and differences in camera field of view, the system does not use a fixed pixel threshold, but rather dynamically calculates it based on the poultry target scale in the second target set. Specifically, the system traverses the second target set that is in an inactive state. Extract the width of the bounding box for each poultry target in the surveillance image data. Next, calculate the arithmetic mean of all bounding box widths. :

[0104] ;

[0105] Where K is the total number of poultry targets in the second target set. The system then multiplies this average by a preset scaling factor. (For example The value ranges from 0.5 to 0.8, resulting in an adaptive matching decision radius. This process ensures that the matching range physically corresponds to a coverage area roughly half the width of a poultry, thereby improving the physical accuracy of the matching.

[0106] After determining the matching radius, the system performs spatial collision detection. The system obtains the coordinates of the geometric center point of the stable repulsion region. and iterate through the second target set. For each inactive individual j, the system calculates its location coordinates. With the center point of the stable repulsion region Euclidean distance between :

[0107] ;

[0108] Subsequently, the system determines whether the Euclidean distance satisfies the preset spatial matching conditions, that is, whether it exists. .

[0109] If the judgment result is true, it indicates that the physical location of the inactive individual j is exactly within the core coverage area of ​​the group's rejection center. This means that the individual has exhibited abnormal movement characteristics of prolonged stillness and constitutes a source of rejection that the group continuously avoids. Based on this, the system determines that the individual is a disease target that triggers abnormal group behavior and immediately generates a health abnormality warning signal for the poultry target. This signal includes the target's unique ID, current location coordinates, and a status tag marked as a high-risk disease / rejection source, and notifies management personnel for timely action through a visual interface highlighting or backend interface push notification.

[0110] Conversely, if the coordinates of no individual in the second target set satisfy the above spatial matching conditions (i.e., the center of the stable rejection region is empty), the system may output an environmental anomaly warning, indicating that there may be non-biological rejection factors (such as foreign objects, water stains, or local environmental mutations) in the region, thereby achieving an effective distinction between biological diseases and environmental anomalies.

[0111] In summary, this embodiment breaks through the limitations of traditional monitoring methods that focus solely on the apparent movement characteristics of individual targets. Instead, it utilizes the mutually exclusive characteristics of group biological behavior as a key basis for disease identification. By mapping static topological gaps in physical space to repulsive force fields at the biodynamic level, and combining morphological diversion and temporal stability verification mechanisms, it successfully isolates stable repulsive regions with clear biological semantics from complex dynamic groups. Furthermore, it uses spatial collision logic to reversely identify abnormal individuals that, although visually appear static, have actually triggered group avoidance responses. This fundamentally solves the problem of the difficulty in binary distinguishing between physiological stillness and pathological stagnation in high-density aquaculture scenarios, significantly improving the detection rate and robustness of health monitoring systems for hidden disease targets.

[0112] Example 2:

[0113] like Figure 3 As shown, the poultry house chicken flock health status behavior image group analysis system includes:

[0114] The image acquisition module is used to acquire monitoring image data of poultry houses and extract features from the monitoring image data to obtain the location coordinates of multiple poultry targets.

[0115] The state analysis module is used to calculate the displacement characteristics of each poultry target based on a preset time window and positioning coordinates, and to divide the poultry targets into a first target set in an active state and a second target set in an inactive state based on the displacement characteristics.

[0116] The topology analysis module is used to construct a discrete topology network based on the positioning coordinates of the first target set, aggregate the grid cells in the discrete topology network that meet the preset spatial sparsity conditions, generate candidate spatial gap regions, and determine the edge targets in the first target set located on the topological boundary of the candidate spatial gap regions.

[0117] The repulsion strength calculation module is used to extract the head orientation vector of the edge target based on the monitoring image data, and generate the bio-repulsion strength index of the candidate spatial gap region according to the radial geometric relationship between the head orientation vector and the center point of the candidate spatial gap region.

[0118] The stability determination module is used to identify candidate spatial void regions whose biological repulsion strength index meets the preset temporal stability conditions as stable repulsion regions.

[0119] The abnormality warning module is used to output a health abnormality warning signal for the corresponding poultry target when the center point of the stable exclusion area and the positioning coordinates of the second target set meet the preset spatial matching conditions.

[0120] The above description is merely an example and illustration of the structure of the present invention. Those skilled in the art can make various modifications or additions to the specific embodiments described, or use similar methods to replace them, as long as they do not deviate from the structure of the invention or exceed the scope defined in the claims, all of which should fall within the protection scope of the present invention.

[0121] In the description of this specification, references to terms such as "an embodiment," "example," "specific example," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0122] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to any specific implementation. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.

Claims

1. A method for analyzing the health status and behavioral patterns of chickens in poultry houses using images, characterized in that... Includes the following steps: Acquire monitoring image data of poultry houses, and extract features from the monitoring image data to obtain the location coordinates of multiple poultry targets; Based on the preset time window and the positioning coordinates, the displacement characteristics of each poultry target are calculated, and based on the displacement characteristics, the poultry targets are divided into a first target set in an active state and a second target set in an inactive state. Based on the positioning coordinates of the first target set, a discrete topology network is constructed, and grid cells in the discrete topology network that meet the preset spatial sparsity condition are aggregated to generate candidate spatial gap regions. The edge targets in the first target set located on the topological boundary of the candidate spatial gap regions are then determined. Based on the monitoring image data, the head orientation vector of the edge target is extracted, and the biorepulsion intensity index of the candidate spatial gap region is generated according to the radial geometric relationship between the head orientation vector and the center point of the candidate spatial gap region. Candidate spatial void regions that satisfy the preset temporal stability condition of the biological repulsion strength index are determined as stable repulsion regions. In response to the fact that the center point of the stable rejection region and the positioning coordinates of the second target set meet the preset spatial matching conditions, a health abnormality warning signal for the corresponding poultry target is output.

2. The method for analyzing the health status behavior of chicken flocks in poultry houses according to claim 1, characterized in that: Based on a preset time window and the positioning coordinates, the displacement characteristics of each poultry target are calculated, and based on the displacement characteristics, the poultry targets are divided into a first target set in an active state and a second target set in an inactive state, including: Extract the sequence of consecutive multi-frame positioning coordinates for each poultry target within the time window; Calculate the Euclidean distance between adjacent frames in the positioning coordinate sequence, and determine the cumulative displacement of the poultry target by summing all the Euclidean distances. Poultry targets whose cumulative displacement is greater than or equal to a preset activity determination threshold are included in the first target set; Poultry targets whose cumulative displacement is less than the preset activity threshold are included in the second target set.

3. The method for analyzing the health status behavior of chicken flocks in poultry houses according to claim 1, characterized in that: The process of generating the candidate spatial gap region includes: The Delaunay triangulation algorithm is used to connect the positioning coordinates in the first target set to generate a triangular mesh diagram composed of multiple triangular mesh units; The area of ​​all triangular grid cells in the triangular grid diagram is counted, the median of all areas is calculated, and the median is multiplied by a preset multiplier to obtain the gap determination area threshold. Triangular mesh cells with an area greater than the void area threshold are identified as mesh cells that satisfy the preset spatial sparseness condition. The connected component labeling algorithm is used to merge spatially adjacent grid cells that meet the preset spatial sparsity condition to obtain the candidate spatial gap region.

4. The method for analyzing the health status behavior of chicken flocks in poultry houses according to claim 1, characterized in that: The process of extracting the head orientation vector includes: Based on the location coordinates of the edge target, a region of interest image containing the edge target is extracted from the monitoring image data; Perform keypoint regression analysis on the region of interest image and output the coordinates of the head keypoint and the center keypoint of the body of the edge target; The difference vector is obtained by subtracting the coordinates of the body center key point from the coordinates of the head key point, and the difference vector is normalized to obtain the head orientation vector of the edge target.

5. The method for analyzing the health status behavior of chicken flocks in poultry houses according to claim 1, characterized in that: The calculation process for the biological rejection strength index includes: Calculate the geometric center coordinates of the candidate spatial gap region; For each edge target, construct a radial centrifugal vector pointing from the geometric center coordinates to the positioning coordinates of that edge target; Calculate the dot product of the head orientation vector of the edge target and the radial centrifugal vector to obtain the centrifugal repulsion factor of the edge target; The biological rejection strength index is generated by weighted aggregation of centrifugal rejection factors of all edge targets based on unilateral suppression logic; the unilateral suppression logic is configured to only accumulate centrifugal rejection factors with positive values ​​and ignore centrifugal rejection factors with negative or zero values.

6. The method for analyzing the health status behavior of chicken flocks in poultry houses according to claim 1, characterized in that: The process of determining the stable repulsion region includes: Obtain the center point of the candidate spatial gap region in the current frame, and the center point of the candidate spatial gap region recorded in the previous frame; Calculate the Euclidean distance between the center points of each candidate spatial gap region in the current frame and the previous frame; Two candidate spatial gap regions with the smallest Euclidean distance and the Euclidean distance being less than the preset tracking radius are identified as the same observation object, and the bio-repulsion intensity index of the same observation object in multiple consecutive frames of monitoring images is extracted to construct the time series of the bio-repulsion intensity index. Calculate the arithmetic mean and variance of the time series; If the arithmetic mean is greater than the preset intensity benchmark and the variance is less than the preset fluctuation benchmark, then the candidate spatial gap region is determined to be the stable repulsion region.

7. The method for analyzing the health status behavior of chicken flocks in poultry houses according to claim 1, characterized in that: In response to the fact that the center point of the stable rejection region and the positioning coordinates in the second target set satisfy a preset spatial matching condition, a health abnormality warning signal for the corresponding poultry target is output, including: Obtain the bounding box width of each poultry target in the second target set in the monitoring image data, calculate the average value of all bounding box widths, and define the preset ratio of the average value as the matching determination radius; Calculate the Euclidean distance between the center point of the stable exclusion region and the location coordinates of any individual in the second target set; If the Euclidean distance is less than the matching determination radius, the preset spatial matching condition is satisfied, and a health abnormality warning signal for the individual is output.

8. The method for analyzing the health status behavior of chicken flocks in poultry houses according to claim 1, characterized in that: After generating the candidate spatial void region, and before generating the biorepulsion strength index of the candidate spatial void region, the method further includes: Calculate the minimum bounding rectangle of the candidate spatial gap region, and extract the major axis length and minor axis length of the minimum bounding rectangle; Calculate the ratio of the major axis length to the minor axis length to obtain the aspect ratio characteristic value of the shape; Candidate spatial void regions whose aspect ratio feature values ​​are less than or equal to a preset morphological threshold are marked as point source repulsion regions, and the step of generating a biological repulsion intensity index for the candidate spatial void regions is further performed for the point source repulsion regions. Candidate spatial gap regions whose aspect ratio features are greater than the preset morphological threshold are marked as suspected linear interference regions, and a linear parallel flow verification step is initiated for the suspected linear interference regions.

9. The method for analyzing the health status behavior of chicken flocks in poultry houses according to claim 8, characterized in that: Initiating a linear parallel flow verification step for the suspected linear interference region includes: Extract the major axis direction vector of the minimum bounding rectangle corresponding to the suspected linear interference region; Based on the preset time window and the positioning coordinates of the edge target, calculate the motion velocity vector of the edge target in the suspected linear interference area at the current moment; Calculate the absolute value of the cosine of the angle between the motion velocity vector of the edge target and the major axis direction vector, and calculate the arithmetic mean of the absolute values ​​of the cosine of the angles corresponding to all edge targets to obtain the parallel flow interference index; If the parallel flow interference index is greater than the preset parallel flow determination threshold, the suspected linear interference region is determined to be a non-biological rejection region, and the biological rejection intensity index of the region is assigned to zero or a preset inhibition value. If the parallel flow interference index is less than or equal to the preset parallel flow determination threshold, the suspected linear interference region is determined to be an effective rejection region, and the step of generating the biological rejection intensity index of the candidate spatial void region is continued for this region.

10. A group analysis system for the health status and behavior of chickens in a poultry house, characterized by: The method for analyzing the health status behavior of chicken flocks in poultry houses according to any one of claims 1-9 includes: The image acquisition module is used to acquire monitoring image data of the poultry house and extract features from the monitoring image data to obtain the positioning coordinates of multiple poultry targets. The state analysis module is used to calculate the displacement characteristics of each poultry target based on a preset time window and the positioning coordinates, and to divide the poultry targets into a first target set in an active state and a second target set in an inactive state based on the displacement characteristics. The topology analysis module is used to construct a discrete topology network based on the positioning coordinates of the first target set, aggregate the grid cells in the discrete topology network that meet the preset spatial sparsity conditions, generate candidate spatial gap regions, and determine the edge targets in the first target set located on the topological boundary of the candidate spatial gap regions. The repulsion strength calculation module is used to extract the head orientation vector of the edge target based on the monitoring image data, and generate the biorepulsion strength index of the candidate spatial gap region according to the radial geometric relationship between the head orientation vector and the center point of the candidate spatial gap region. The stability determination module is used to determine the candidate spatial gap regions that meet the preset temporal stability conditions of the biological repulsion strength index as stable repulsion regions. An abnormality warning module is used to output a health abnormality warning signal for the corresponding poultry target when the center point of the stable rejection area and the positioning coordinates of the second target set meet a preset spatial matching condition.

Images