Intrusion alerting method and system for a rearing shed

Abstract

The application relates to the technical field of alarm, and provides an invasion alarm method and system for a breeding house, multi-source acquisition and space-time registration fusion of point cloud data, thermal imaging signals and vibration signals are realized through a sensing module, and the accuracy and stability of sensing data are improved; a parsing module extracts behavior characteristics of different kinds of poultry through clustering analysis based on historical sensing data, and constructs a space-time behavior baseline corresponding to a legal target; a recognition module compares multi-dimensional characteristics with the space-time behavior baseline through multi-dimensional characteristic extraction and hierarchical matching, and can quickly and accurately determine whether a target is a legal target or an invasion target; a response module adopts a hierarchical response mechanism, first executes a driving action on the invasion target, and only sends alarm information and links the physical defense facilities when the invasion target continuously exists or a damage event is triggered, so that the timeliness and effectiveness of invasion prevention and control are ensured, and the stress reaction of poultry caused by meaningless alarm and defense linkage is avoided.

Description

Technical Field

[0001] This application relates to the field of alarm technology, and in particular to intrusion alarm methods and systems for livestock sheds. Background Technology

[0002] Currently, intrusion alarms in livestock sheds have gradually evolved from single-sensor detection to multi-source monitoring. The industry generally adopts single or combined detection methods such as infrared sensing, simple video monitoring, and vibration detection. Alarms are triggered by monitoring moving targets in the livestock shed. Some technologies have achieved basic sensor signal acquisition and on-site alarm functions, which to a certain extent meet the basic security needs of livestock sheds.

[0003] However, existing intrusion alarm technologies for livestock sheds have certain shortcomings in practical applications. They mostly use independent detection methods such as video or vibration, which are easily affected by environmental interference such as dust, changes in light, and normal activities of poultry in the livestock shed. The accuracy of target detection is low. It is difficult to identify the behavioral characteristics of legitimate targets in the livestock shed, which may lead to misjudging legitimate activities as intrusion. Frequent alarms can interfere with normal farming operations and may also cause staff to become desensitized to alarm signals, thus reducing the effectiveness of prevention and control.

[0004] Based on the shortcomings of the existing technology, the technical problem to be solved by this application is how to achieve accurate intrusion detection and alarm in breeding sheds in complex environmental scenarios. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this application provides an intrusion alarm method and system for livestock sheds.

[0006] In the first aspect, this application provides an intrusion alarm system for livestock sheds, which includes: a sensing module, a parsing module, an identification module, and a response module;

[0007] The sensing module is used to acquire point cloud data and thermal imaging signals of moving targets in the breeding house, as well as vibration signals at the inlet and outlet channels. After performing spatiotemporal registration, it is fused to generate sensing data including point cloud features, thermal radiation gradient and vibration spectrum.

[0008] The parsing module is used to extract and distinguish the activity areas, body shape characteristics and gait frequency characteristics of different types of poultry based on historically acquired perception data through cluster analysis, and to construct a spatiotemporal behavioral baseline that includes behavioral baseline characteristics and dynamic characteristics.

[0009] The identification module is used to extract moving targets from the current perception data and obtain the multi-dimensional features of the moving targets. The multi-dimensional features are then matched hierarchically with the spatiotemporal behavior baseline. If the match is successful, the moving target is determined to be a legitimate target; if the match fails, the moving target is determined to be an intrusion target.

[0010] The response module is used to execute a graded response when the moving target is determined to be an intrusion target or when an abnormal cloud of moving target points is detected. When the intrusion target persists or a preset destructive event is triggered, alarm information is sent and physical defense facilities are activated.

[0011] As an optional implementation, the sensing data fused and generated in the sensing module is specifically used for:

[0012] Acquire point cloud data, thermal imaging signals, and vibration signals at the inlet and outlet channels of moving targets within the breeding shed; and perform spatiotemporal registration of the point cloud data and thermal imaging signals based on the vibration signals.

[0013] Using the moving target as a dynamic background, the system identifies the point cloud clusters of the moving target from the registered point cloud data and extracts the features of the point cloud clusters. It also extracts the thermal radiation gradient of the moving target from the registered thermal imaging signal and converts the vibration signal into the corresponding vibration spectrum.

[0014] Using vibration spectrum as a link, the point cloud features of the same moving target are bound to thermal radiation gradient and fused to generate sensing data.

[0015] As an optional implementation, the spatiotemporal registration performed in the sensing module is specifically used for:

[0016] Time-domain peak detection is performed on vibration signals to screen vibration signals triggered by moving targets and determine vibration time-domain anchor points;

[0017] Using the vibration time-domain anchor point as a reference, point cloud data frames and thermal imaging signal frames corresponding to the time window are extracted;

[0018] Based on the propagation attenuation parameters of the vibration signal, the spatial coordinates of the moving target in the inlet and outlet channels of the breeding house are deduced.

[0019] Using the spatial coordinates as the registration center, local spatial registration is performed on the point cloud data frame and the thermal imaging signal frame to complete the spatiotemporal registration.

[0020] As an optional implementation, the cluster analysis in the parsing module is specifically used for:

[0021] The historically acquired sensory data is preprocessed, and invalid samples are removed based on the preset thermal radiation gradient threshold and gait frequency threshold in the poultry house, while candidate samples of poultry are retained.

[0022] Using vibration spectrum and thermal radiation gradient as weighted features, density clustering was performed on candidate samples to separate different types of poultry samples.

[0023] The body shape parameters, spatial coordinates, and gait frequency of the vibration spectrum of the corresponding point cloud features of poultry samples are extracted. After similarity clustering is performed, the activity area, body shape characteristics, and gait frequency characteristics of different types of poultry are obtained by combining the spatial zoning constraints of the breeding house.

[0024] Based on the surface thermal radiation data of moving targets under different ambient temperatures, the thermal radiation gradient is fitted and dynamic characteristics of different types of poultry adapted to ambient temperature are constructed.

[0025] As an optional implementation, the similarity clustering performed in the parsing module is specifically used for:

[0026] Calculate the similarity of body shape parameters and spatial coordinate distance of corresponding point cloud features in poultry samples, and cluster the poultry samples using Euclidean distance;

[0027] Based on the spatial zoning constraints of the poultry house, the activity areas and body size characteristics of different types of poultry are matched;

[0028] Gait frequencies corresponding to poultry samples are extracted, and the poultry samples are clustered using a density clustering algorithm. The consistency of gait frequencies in different clustering results is compared to obtain the gait frequency characteristics of different types of poultry.

[0029] As an optional implementation, the construction of the spatiotemporal behavioral baseline in the parsing module is specifically used for:

[0030] The activity area, body size characteristics, and gait frequency characteristics of different types of poultry were set as the behavioral benchmark characteristics of poultry.

[0031] By binding the gait frequency characteristics of the same type of poultry with the corresponding activity area and body size characteristics, spatiotemporal association pairs are formed for each type of poultry.

[0032] Based on historically acquired perception data, dynamic threshold ranges are set for the spatiotemporal correlation of each type of poultry, corresponding behavioral baseline features and dynamic features.

[0033] For each type of poultry, a sub-baseline is established, which includes the corresponding behavioral baseline features, the dynamic features, the dynamic threshold range, and the spatiotemporal association rules. All sub-baselines are integrated to form a spatiotemporal behavioral baseline.

[0034] As an optional implementation, the multidimensional features of the moving target acquired in the recognition module are specifically used for:

[0035] The current sensing data is preprocessed, and the core area of ​​the moving target is located based on the feature of the point cloud, with the moving target as the dynamic background. The thermal radiation gradient and vibration spectrum corresponding to the core area are then locked.

[0036] Gait frequency features are extracted from the locked vibration spectrum, body shape parameters are extracted from the point cloud features, and gradient values ​​adapted to the current ambient temperature are extracted from the thermal radiation gradient as core features.

[0037] Extract the state of the moving target point cloud and the real-time spatial coordinates of the moving target from the current sensing data as auxiliary features;

[0038] The core features and auxiliary features are verified to determine whether they belong to the same moving target. Based on the gait frequency features, the verified core features and auxiliary features are correlated and normalized to form a multidimensional feature of the moving target.

[0039] As an optional implementation, the hierarchical matching is used for:

[0040] The core features of the multidimensional features are compared with the behavioral benchmark features of the corresponding sub-baseline dimension by dimension to determine whether the core features are within the dynamic threshold range.

[0041] If the core feature is not within the dynamic threshold range, the current sub-baseline is directly determined to be a failure to match, and the matching of other types of poultry sub-baselines is switched.

[0042] If the core feature is within the dynamic threshold range, the auxiliary features of the multidimensional feature are compared with the behavioral baseline features, spatiotemporal association rules, and the dynamic features of the corresponding sub-baseline.

[0043] If the auxiliary feature comparison is successful, then verify whether the multidimensional features involved in the matching are valid features of the same moving target;

[0044] If the verification passes, the match is considered successful and the corresponding moving target is determined to be a legitimate target. If the verification fails or the auxiliary feature comparison is unqualified, the corresponding sub-baseline is considered to have failed to match. If all sub-baselines fail to match, the moving target is considered to be an intrusion target.

[0045] As an optional implementation, the response module is used for:

[0046] When a moving target is determined to be an intrusion target or an anomaly is detected in the point cloud of a moving target, a graded response is immediately executed to carry out corresponding expulsion actions, and the existence status of the intrusion target and whether a preset destruction event is triggered are monitored in real time.

[0047] If the intrusion target persists or a preset sabotage event is triggered, a corresponding alarm message will be generated and sent out, and the linkage defense action of the physical defense facilities will be activated at the same time.

[0048] If the intrusion target disappears after being driven away and no preset destruction event is triggered, the driving action will automatically stop and no alarm information will be sent or physical defense facilities will be linked.

[0049] Secondly, this application provides an intrusion alarm method for aquaculture sheds, which includes: acquiring point cloud data and thermal imaging signals of moving targets in aquaculture sheds, as well as vibration signals at entrance and exit channels, and performing spatiotemporal registration to fuse and generate sensing data including point cloud features, thermal radiation gradient and vibration spectrum.

[0050] Based on historically acquired sensory data, cluster analysis is used to extract and distinguish the activity areas, body size characteristics, and gait frequency characteristics of different types of poultry, so as to construct a spatiotemporal behavioral baseline that includes behavioral baseline characteristics and dynamic characteristics.

[0051] Extract moving targets from the current perception data and obtain multi-dimensional features of the moving targets. Perform hierarchical matching of the multi-dimensional features with the spatiotemporal behavior baseline. If the matching is successful, the moving target is determined to be a legitimate target. If the matching fails, the moving target is determined to be an intrusion target.

[0052] When a moving target is determined to be an intrusion target or an abnormal cloud of moving target points is detected, a tiered response is executed. If the intrusion target persists or a preset destructive event is triggered, an alarm message is sent and physical defense facilities are activated.

[0053] Compared with existing technologies, the beneficial effects of this application are as follows: By using a sensing module to acquire and spatiotemporally register and fuse point cloud data, thermal imaging signals, and vibration signals within the poultry house from multiple sources, the defects of single-modal sensing being susceptible to environmental interference are avoided, thus improving the integrity and anti-interference capability of the sensing data. Through a parsing module, a poultry spatiotemporal behavior baseline is constructed based on cluster analysis of historical sensing data, establishing benchmark characteristics for normal poultry activities and providing a reference for target legitimacy identification. Through a recognition module, multi-dimensional features of moving targets are hierarchically matched with the spatiotemporal behavior baseline, accurately distinguishing legitimate targets from intrusive targets within the poultry house, achieving the detection of point cloud anomalies, and reducing the target misjudgment rate. Through a response module, graded responses are executed for intrusive targets or point cloud anomalies, and alarms and physical defense linkages are triggered in conjunction with the intrusion target status and destructive events, achieving differentiated and precise security measures for the poultry house, and improving the intelligence level and security reliability of the poultry house intrusion alarm system. Attached Figure Description

[0054] To more clearly illustrate the technical solutions of the embodiments of this application, the drawings used in the description of the embodiments will be briefly introduced below. The drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained from these drawings without creative effort. Wherein:

[0055] Figure 1 This is a system flowchart of an intrusion alarm system for livestock sheds provided in an embodiment of this application;

[0056] Figure 2A logical flowchart for constructing a spatiotemporal behavioral baseline provided in an embodiment of this application;

[0057] Figure 3 This is a flowchart illustrating an intrusion alarm method for livestock sheds provided in an embodiment of this application. Detailed Implementation

[0058] To make the objectives, technical solutions, and advantages of the embodiments of this application more apparent and understandable, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments.

[0059] Example 1:

[0060] like Figure 1 The diagram shown is a system flowchart of an intrusion alarm system for livestock sheds, which is provided for an embodiment of this application. The system includes a sensing module, a parsing module, an identification module, and a response module.

[0061] The sensing module is used to acquire point cloud data and thermal imaging signals of moving targets in the breeding house, as well as vibration signals at the inlet and outlet channels. After performing spatiotemporal registration, the data is fused to generate sensing data including point cloud features, thermal radiation gradient and vibration spectrum.

[0062] Furthermore, the fused and generated sensory data is specifically used for:

[0063] Acquire point cloud data, thermal imaging signals and vibration signals at the inlet and outlet channels of moving targets inside the breeding house, and perform spatiotemporal registration of point cloud data and thermal imaging signals based on vibration signals;

[0064] Using the moving target as a dynamic background, the system identifies the point cloud clusters of the moving target from the registered point cloud data and extracts the features of the point cloud clusters. It also extracts the thermal radiation gradient of the moving target from the registered thermal imaging signal and converts the vibration signal into the corresponding vibration spectrum.

[0065] Using vibration spectrum as a link, the point cloud features of the same moving target are bound to thermal radiation gradient and fused to generate sensing data.

[0066] Point cloud data, thermal imaging signals, and vibration signals within the breeding shed are acquired independently by different types of sensing devices. LiDAR and infrared thermal imagers are installed inside the breeding shed to acquire data on moving targets, while vibration sensors are only deployed on the rigid structure of the breeding shed's entrance and exit passages. These three types of devices have objective problems such as physical installation position deviations and asynchronous data acquisition timing. Uncalibrated heterogeneous data will directly lead to misalignment in the extraction of moving target features. Vibration signals are collected by the rigid structure of the entrance and exit passages. Vibration signals triggered by moving targets have strong real-time performance and precise triggering positions. They are not affected by the bedding material or soil medium on the breeding shed floor and have a natural advantage as a unified spatiotemporal reference.

[0067] Acquiring point cloud data refers to obtaining three-dimensional spatial point set data corresponding to moving targets through lidar deployed inside the breeding shed; acquiring thermal imaging signals refers to obtaining infrared thermal radiation data of moving targets through infrared thermal imagers; acquiring vibration signals at the entrance and exit passages refers to deploying vibration sensors only on rigid structures such as metal door frames and fences at the entrances and exits of the breeding sheds to obtain vibration signals triggered by moving targets passing through the passages; performing spatiotemporal registration based on vibration signals involves using the trigger time of the vibration signal as the time reference and the corresponding entrance and exit spatial position as the spatial reference to perform temporal and spatial coordinate calibration on the point cloud data acquired by lidar and the thermal imaging signals acquired by infrared thermal imagers, so that the three types of signals correspond to the same spatiotemporal state of the same moving target.

[0068] The effective vibration trigger threshold is a preset value for filtering effective vibration signals, set to 1.5 times the vibration amplitude of the static environment in the livestock shed, to eliminate invalid interference such as environmental disturbances and equipment resonance. For example, if the vibration amplitude of the static environment in the livestock shed is 0.33g, then the effective vibration trigger threshold is set to 0.5g. For instance, vibration sensors are installed on the rigid structure of the metal door frame at the entrance and exit of the livestock shed, with an effective vibration trigger threshold of 0.5g. The lidar is fixed to the ceiling inside the shed, and the spatial range of the livestock shed for acquiring point cloud data is 0m to 10m in length and 0m to 5m in width. m, height 0m~2.5m; infrared thermal imager temperature measurement range 0℃~50℃, frame rate 25fps; when poultry such as chickens, ducks and geese pass through the entrance and exit channels, or when intruding targets such as weasels enter the entrance and exit, the amplitude of the triggered vibration signal exceeds 0.5g, which is judged as a valid trigger; using the trigger time as the time reference and the center coordinates of the entrance and exit (5m, 2.5m, 0m) as the spatial reference, the acquisition timing and spatial coordinates of the lidar and infrared thermal imager are automatically calibrated so that the point cloud data, thermal imaging signal and vibration signal correspond to the same moving target.

[0069] This allows for the spatiotemporal calibration of multi-source data using vibration signal benchmarks, unifying the spatiotemporal coordinate system of three types of heterogeneous data, avoiding interference from the ground medium of the livestock shed on vibration monitoring, achieving source matching of point cloud data, thermal imaging signals and vibration signals, eliminating feature inaccuracies caused by spatiotemporal deviations from the source of data acquisition, and ensuring the accuracy of subsequent feature extraction.

[0070] The original data after spatiotemporal registration contains redundant static environmental information such as the walls of the poultry house, feeding equipment, and poultry house supports. This redundant information can interfere with the accurate extraction of moving target features. At the same time, the original vibration signal is a time-domain electrical signal, which cannot be directly used for subsequent clustering analysis and feature matching. The original thermal imaging data can only reflect temperature values ​​and cannot reflect the difference in thermal radiation between the moving target and the environment. By using the moving target as a dynamic background, eliminating static environmental interference, focusing on the moving target itself to extract key features, converting the vibration signal into a vibration spectrum, and the thermal imaging signal into a thermal radiation gradient, the original data is transformed into quantitative features that are easy to analyze later, which is suitable for the identification needs of distinguishing poultry from intrusive targets in poultry houses.

[0071] Using a moving target as a dynamic background means removing the static environment within the breeding shed as background and focusing only on the area where the moving target is located for feature processing; moving target point cloud refers to a continuous set of three-dimensional point clouds belonging to the same moving target in the registered point cloud data, which is different from the static environment point cloud; point cloud features refer to the body shape and contour-related features extracted from the moving target point cloud to characterize the target's morphology; thermal radiation gradient refers to the difference gradient between the thermal radiation on the surface of the moving target and the thermal radiation of the breeding shed environment extracted from the registered thermal imaging signal, rather than a single temperature value; vibration spectrum refers to the frequency domain features characterizing the vibration characteristics of the moving target obtained after processing the time-domain vibration signal through frequency domain transformation.

[0072] For example, static background removal processing is performed on the registered point cloud data to remove point clouds of fixed facilities such as walls and supports. Using moving targets as the dynamic background, the size range of point cloud clusters corresponding to chickens, ducks, and geese is identified as 0.3m-0.8m in length, 0.2m-0.4m in width, and 0.2m-0.6m in height, while the size range of point cloud clusters corresponding to weasels is 0.25m-0.45m in length, 0.12m-0.2m in width, and 0.15m-0.3m in height. From the registered thermal imaging signal, the thermal radiation difference between the surface of the moving target and the environment is calculated to form a thermal radiation gradient. The surface temperature of poultry is 38℃-42℃, and the ambient temperature is 20℃-25℃, with a thermal radiation gradient difference of 13℃-22℃. The time-domain vibration signal is converted into a vibration spectrum. The main frequency range of the trigger spectrum for poultry movement is 10Hz-30Hz, and the main frequency range of the trigger spectrum for weasel intrusion is 35Hz-60Hz, in order to distinguish the vibration characteristics of different targets.

[0073] This process eliminates redundancy and refines features from the original data, retaining only the point cloud features, thermal radiation gradients, and vibration spectra specific to moving targets, while removing static environmental interference. At the same time, it transforms the original time-domain signals and thermal imaging values ​​into quantitative features adapted to recognition needs, highlighting the differences in characteristics between poultry such as chickens, ducks, and geese and intruding targets such as weasels. This provides support for distinguishing between legitimate and intruding targets and avoids interference from redundant features in the generation of fused sensing data.

[0074] The features of moving target point cloud, thermal radiation gradient, and vibration spectrum are heterogeneous features of different dimensions. The dispersed storage of heterogeneous features is prone to confusion in the scenario of multiple targets coexisting in a livestock house, especially when poultry such as chickens, ducks, and geese appear at the same time as intrusive targets such as weasels, which can easily lead to feature overlap and misalignment. The vibration spectrum is triggered by the moving target, and it can be used as a link to accurately bind the three types of heterogeneous features, clarify the relationship between the features and the moving target, and form integrated fusion data belonging to the same moving target. This ensures that the subsequent analysis and recognition modules can accurately call the feature data of the corresponding target.

[0075] Using vibration spectrum as a link means taking advantage of the uniqueness of the vibration spectrum triggered by a single moving target, using it as a unique identifier for feature association, and assigning the same identifier to the corresponding extracted point cloud features and thermal radiation gradient, thus clarifying the subject to which the three types of features belong; binding features of the same moving target means linking and integrating point cloud features with thermal radiation gradients that have the same association identifier; fusing to generate sensing data means integrating the three types of bound features into an integrated data unit to form sensing data.

[0076] For example, when chickens pass through the entrance / exit passage in the poultry house, the main frequency of the vibration spectrum is 15Hz-25Hz. This spectrum is used as an association identifier, and the same identifier is assigned to the corresponding point cloud cluster with a length of 0.4m, a width of 0.25m, a height of 0.3m, and a thermal radiation gradient difference of 18℃, clearly indicating that the three types of features belong to the same chicken flock target. If a weasel enters, the main frequency of the vibration spectrum is 40Hz-50Hz. This is used as an identifier to bind the point cloud cluster with a length of 0.3m, a width of 0.15m, a height of 0.2m, and a thermal radiation gradient difference of 16℃, generating the corresponding intrusion target perception data. When multiple moving targets exist in the poultry house at the same time, the corresponding features are bound separately by the numerical differences in the vibration spectrum 10Hz-30Hz (poultry) and 35Hz-60Hz (weasel), generating multiple independent perception data, preventing the confusion of multiple target features.

[0077] This enables the integrated integration of heterogeneous features, clarifies the attribution relationship between features and moving targets, ensures the integrity and uniqueness of fused perception data, solves the feature confusion problem in multi-target scenarios in livestock sheds, accurately distinguishes the feature data of legitimate poultry such as chickens, ducks and geese from intruding targets such as weasels, and provides a standardized and homogeneous core data carrier for subsequent module processing. It is the core data prerequisite for the subsequent operation of the entire intrusion alarm system.

[0078] Specifically, spatiotemporal registration is performed for:

[0079] Time-domain peak detection is performed on vibration signals to screen vibration signals triggered by moving targets and determine vibration time-domain anchor points;

[0080] Using the vibration time-domain anchor point as a reference, point cloud data frames and thermal imaging signal frames corresponding to the time window are extracted;

[0081] Based on the propagation attenuation parameters of the vibration signal, the spatial coordinates of the moving target in the inlet and outlet channels of the breeding house are deduced.

[0082] Using spatial coordinates as the registration center, local spatial registration is performed on point cloud data frames and thermal imaging signal frames to complete spatiotemporal registration.

[0083] Vibration signals acquired from the rigid structure of the entrance and exit passage of the livestock shed are not all valid signals triggered by moving targets. They include invalid interference vibrations such as wind blowing the fence, equipment resonance, and environmental disturbances. Such interference signals can cause deviations in the time reference of spatiotemporal registration. Time domain peak detection can accurately filter out the valid vibration signals triggered by poultry such as chickens, ducks, and geese passing through, or intruding targets such as weasels entering from the mixed vibration signals, so as to determine a unique and accurate vibration time domain anchor point, providing a stable reference for subsequent time dimension registration and avoiding registration errors caused by interference signals from the source.

[0084] Time-domain peak detection refers to scanning the time-domain waveform of a vibration signal segment by segment to extract the peak point of the signal amplitude; vibration signal triggered by a moving target refers to an effective vibration signal whose amplitude exceeds the effective vibration trigger threshold and is generated by poultry or intruding targets contacting the rigid structure of the inlet / outlet; vibration time-domain anchor point refers to the moment corresponding to the peak of the effective vibration signal, which is the only time reference for subsequent data frame interception.

[0085] For example, vibration sensors are installed at the entrance and exit passages of the poultry house. The vibration detection range of the sensors is set to 0g to 2g, and the effective vibration trigger threshold is 0.5g. A sliding window with a duration of 0.05 seconds is used to scan the vibration signal window by window. When the flock of chickens passes through the entrance and exit passages, the amplitude of the triggered vibration signal reaches 0.8g, which exceeds the preset threshold. The signal is determined to be an effective vibration signal triggered by a moving target, and the time corresponding to the peak value is determined as the vibration time domain anchor point. If the amplitude of the vibration signal generated by environmental disturbance is only 0.2g, which does not exceed the threshold, the invalid signal is discarded, and no vibration time domain anchor point is generated.

[0086] By effectively eliminating environmental interference vibrations through time-domain peak detection, the system accurately selects valid vibration signals triggered by moving targets, determines a unique and stable vibration time-domain anchor point, provides an accurate time reference for subsequent time window interception, and avoids timing deviation problems caused by invalid vibration signals.

[0087] LiDAR and infrared thermal imagers are continuous acquisition devices that continuously output a large number of data frames. If all data frames are used directly, a large amount of redundant data unrelated to the moving target will be included, which will significantly increase the computational load of subsequent spatial registration. At the same time, it cannot be guaranteed that the data frames correspond perfectly with the moment when the moving target triggers vibration. By using the vibration time domain anchor point as a reference to extract data frames within a fixed time window, the point cloud data frames and thermal imaging signal frames that are spatiotemporally synchronized with the moving target can be accurately screened out, and redundant data can be eliminated to ensure the relevance of the data used for subsequent registration.

[0088] The time window refers to a fixed time range extending forward and backward from the vibration time domain anchor point. It is used to limit the range of data frames to be captured. The fixed duration is set according to the sampling frame rate of the lidar and infrared thermal imager to ensure that the time window fully covers at least one valid data frame. Capturing the data frame corresponding to the time window means selecting the point cloud data frames and thermal imaging signal frames whose timestamps fall within the time window as the raw data for subsequent registration.

[0089] For example, the sampling frame rate of the lidar is 10 frames / second, with a single frame duration of 0.1 seconds, and the sampling frame rate of the infrared thermal imager is 25 frames / second, with a single frame duration of 0.04 seconds. Based on this, a time window is set with the vibration time domain anchor point as the center, extending 0.1 seconds before and after it, for a total duration of 0.2 seconds. The acquisition frame rate of the lidar is 10 frames / second, and the acquisition frame rate of the infrared thermal imager is 25 frames / second. When extracting data according to this time window, 2 point cloud data frames and 5 thermal imaging signal frames within this time period are selected. The timestamps of all extracted data frames are highly matched with the vibration time domain anchor point, corresponding to the trigger time of the same moving target.

[0090] This enables precise selection of point cloud data frames and thermal imaging signal frames in the time dimension, eliminating redundant data frames unrelated to the moving target, reducing the computational load of subsequent spatial registration, and ensuring that the two types of data frames are completely synchronized with the vibration signal in time, providing time-unified raw data for subsequent spatial coordinate inversion and local registration.

[0091] Spatiotemporal registration requires not only synchronization in the time dimension, but also the determination of the precise spatial location of the moving target as the registration center. The inlet and outlet channels of the breeding house are rigid structures. When the vibration signal propagates in the rigid structure, its amplitude will show a fixed attenuation law with the propagation distance. Using this propagation attenuation parameter, the spatial coordinates of the moving target in the inlet and outlet channels can be accurately deduced, providing a stable spatial reference for subsequent local spatial registration.

[0092] The propagation attenuation parameter of the vibration signal refers to the fixed coefficient by which the amplitude of the vibration signal attenuates with the propagation distance when it propagates in the rigid structure of the inlet and outlet. It is obtained by calibration before the system is deployed. The reverse spatial coordinate refers to the calculation of the propagation distance by combining the effective vibration signal amplitude collected by at least two vibration sensors with the propagation attenuation parameter, and then obtaining the three-dimensional spatial coordinate of the moving target through the positioning algorithm. The coordinate is limited to the range of the inlet and outlet passage of the breeding house.

[0093] Vibration signals in rigid structures follow a logarithmic linear decay law, and their amplitude decay calculation formula is: L = L0 - α × d, where L is the vibration amplitude acquired by the sensor, L0 is the initial vibration amplitude triggered by the moving target, α is the vibration propagation attenuation coefficient, and d is the propagation distance from the moving target to the vibration sensor. Since the two sensors acquire the same vibration source signal, the initial vibration amplitude L0 is the same. Solving the attenuation equations of the two sensors simultaneously can eliminate L0, and the distance from the moving target to the two sensors can be directly calculated. Then, combined with the known coordinates of the sensors, the spatial coordinates of the moving target can be calculated through a linear positioning algorithm.

[0094] For example, two vibration sensors are installed on the rigid structure of the entrance and exit passage of a livestock shed. The coordinates of the left sensor are (4.5m, 2.5m, 0m), and the coordinates of the right sensor are (5.5m, 2.5m, 0m). The distance between the two sensors is 1m, and the pre-calibrated vibration propagation attenuation coefficient is 0.02dB / m. After the moving target triggers the vibration, the vibration amplitude obtained by the left sensor is 0.8g, and the vibration amplitude obtained by the right sensor is 0.78g. Solving the simultaneous attenuation equations: 0.8 = L0 - 0.02 × d1, 0.78 = L0 - 0.02 × d2, subtracting the two equations to eliminate L0, we get 0.02 = 0.02 × (d2 - d1). Given the constraints of a straight, rigid structure at the entrance / exit channel and a 1m distance between the two sensors, the distance d1 from the moving target to the left sensor is calculated to be 1m, and the distance d2 from the moving target to the right sensor is 0m. Based on the X-axis coordinates of 4.5m for the left sensor and 5.5m for the right sensor, the distance from the moving target to the right sensor is 0m, meaning the moving target and the right sensor are at the same X-axis position. The Y-axis coordinate is the same as the sensor, at 2.5m. Since both poultry and the intruding target move close to the ground, the Z-axis coordinate is taken as 0.1m. Finally, the spatial coordinates of the moving target in the entrance / exit channel are obtained by reverse calculation using a linear positioning algorithm: (5.5m, 2.5m, 0.1m).

[0095] Based on the propagation and attenuation law of vibration signals in rigid structures, the spatial coordinates of the moving target in the inlet and outlet channels can be accurately deduced, solving the problem of spatial positioning deviation between point cloud data and thermal imaging signals. This provides a unique and accurate spatial registration center for subsequent local spatial registration, determining the effect of local registration between point cloud data frames and thermal imaging signal frames.

[0096] Even after time-synchronized point cloud data frames and thermal imaging signal frames, spatial offsets still exist due to differences in the installation positions of the lidar and infrared thermal imager and equipment calibration errors. Without spatial registration, the spatial positions of the moving targets corresponding to the two types of signals cannot coincide. Performing local spatial registration with the spatial coordinates of the moving target as the registration center can focus on the local area where the moving target is located for calibration, avoiding redundant calculations in global registration and eliminating spatial offsets between the two types of data.

[0097] The registration center refers to the spatial coordinates of the moving target obtained by reverse calculation, which is the core reference point for local spatial registration. Local spatial registration refers to calibrating the spatial coordinates of point cloud data frames and thermal imaging signal frames only with the registration center as the core and limiting the local spatial range, rather than global registration. Completing spatiotemporal registration means achieving complete unification of point cloud data, thermal imaging signals, and vibration signals in the time and space dimensions, corresponding to the same moving target.

[0098] The feature matching and coordinate adjustment of the iterative nearest point algorithm are as follows: extract the outline centroid features of the moving target in the point cloud data, extract the temperature centroid features of the corresponding area in the thermal imaging signal, and use the two centroids as matching feature points; calculate the translation vector and rotation angle between the two sets of feature points, reduce the distance error between feature points through iterative optimization, and adjust the spatial coordinates of the point cloud data by translation and rotation according to the optimized translation vector and rotation angle until the centroid of the point cloud outline and the temperature centroid of the thermal imaging are completely coincident, thus completing the spatial calibration; the parameters of the iterative nearest point algorithm are set to an upper limit of 20 iterations, a convergence threshold of 0.01m, and a nearest point search radius of 0.5m. Those skilled in the art can adjust the parameters in practice.

[0099] Specifically, first extract the centroid features (X) of the moving target from the point cloud data. p ,Y p Z p Extract the temperature centroid features (X) of the region corresponding to the thermal imaging signal. h ,Y h Z h The two types of centroids are used as matching feature points; with a search radius of 0.5m, the closest corresponding points between the centroids of the point cloud and the centroids of the thermal imaging are matched; based on the two sets of feature points, the translation vector (ΔX, ΔY, ΔZ) = (X... h -X p ,Y h -Y p Z h -Z p Construct a coordinate transformation matrix using the coordinate transformation matrix and rotation angles (α, β, γ); perform translation and rotation adjustments on the point cloud data based on the coordinate transformation matrix, and calculate the distance error of the adjusted feature points; if the error is ≤0.01m (convergence threshold) or reaches the upper limit of 20 iterations, stop the iteration; the centroid of the point cloud contour coincides with the centroid of the thermal imaging temperature, and the spatial offset is eliminated.

[0100] For example, using the back-derived (5.5m, 2.5m, 0.1m) as the registration center, a local registration region with a side length of 1m is defined, covering the range of 4.5m to 5.5m, 2m to 3m, and 0m to 1m. Point cloud data and thermal imaging signal data within this local region are extracted, and feature matching and coordinate adjustment are performed using the iterative nearest point algorithm. The spatial coordinates of the point cloud data and the spatial coordinates of the thermal imaging signal are calibrated to be completely coincident, eliminating the spatial offset between the two types of data and completing the spatiotemporal registration.

[0101] This allows for precise elimination of spatial offset between point cloud data frames and thermal imaging signal frames through local spatial registration. Combined with time synchronization in the preceding steps, it achieves complete unification of the three types of signals in the spatiotemporal dimensions, ensuring that all data correspond to the same moving target. This provides accurate and homogeneous basic data for subsequent fusion and generation of sensing data, determining the accuracy of subsequent point cloud feature extraction and thermal radiation gradient acquisition. It is the key foundation for the entire sensing module's data processing.

[0102] The parsing module is used to extract and distinguish the activity areas, body shape characteristics and gait frequency characteristics of different types of poultry based on historically acquired perception data through cluster analysis, and to construct a spatiotemporal behavioral baseline that includes behavioral baseline characteristics and dynamic characteristics.

[0103] Furthermore, cluster analysis is used for:

[0104] The historically acquired sensory data is preprocessed, and invalid samples are removed based on the preset thermal radiation gradient threshold and gait frequency threshold in the poultry house, while candidate samples of poultry are retained.

[0105] Using vibration spectrum and thermal radiation gradient as weighted features, density clustering was performed on candidate samples to separate different types of poultry samples.

[0106] The body shape parameters, spatial coordinates, and gait frequency of the vibration spectrum of the corresponding point cloud features of poultry samples are extracted. After similarity clustering is performed, the activity area, body shape characteristics, and gait frequency characteristics of different types of poultry are obtained by combining the spatial zoning constraints of the breeding house.

[0107] Based on the surface thermal radiation data of moving targets under different ambient temperatures, the thermal radiation gradient is fitted and dynamic characteristics of different types of poultry adapted to ambient temperature are constructed.

[0108] Historical sensing data is continuously acquired and stored by the sensing module. The data includes valid samples of poultry such as chickens, ducks, and geese, as well as invalid samples corresponding to environmental interference, equipment malfunctions, and intrusion targets such as weasels. Invalid samples directly interfere with the accuracy of cluster analysis, leading to deviations in subsequent feature extraction. The preset thermal radiation gradient threshold and gait frequency threshold are screening criteria set based on the inherent physiological and movement characteristics of poultry. By filtering through the thresholds, invalid samples can be eliminated, and only candidate samples of legitimate poultry can be retained, providing a clean data foundation for subsequent cluster analysis.

[0109] Historically acquired sensing data refers to historically fused data that has completed spatiotemporal registration and includes point cloud features, thermal radiation gradients, and vibration spectra. Preprocessing refers to removing random noise from the sensing data and regularizing samples with missing features using filtering algorithms. The thermal radiation gradient threshold is the range of thermal radiation gradient values ​​that distinguishes poultry from invalid targets. By measuring the difference in thermal radiation between the body surface of chickens, ducks, and geese and the environment of the breeding house at different ambient temperatures (10℃-30℃), the effective range of 95% of legitimate poultry samples is taken, excluding the thermal radiation gradient range of intrusive targets such as weasels, thus accurately distinguishing legitimate poultry from intrusive targets and environmental interference. The gait frequency threshold is the range of vibration frequencies that distinguishes normal poultry movement from abnormal interference. By measuring the vibration spectrum frequencies corresponding to poultry walking, normal walking, and sprinting, the frequency range of normal poultry movement is obtained, excluding the abnormal frequency ranges of intrusive target movement, equipment vibration, and environmental disturbance. Invalid samples refer to environmental interference data and intrusive target data that do not meet the threshold requirements. Poultry candidate samples refer to the sensing data corresponding to chickens, ducks, and geese that meet the threshold requirements.

[0110] For example, according to actual measurements, when the ambient temperature is 20℃, the thermal radiation difference between the body surface of chickens, ducks, and geese and the environment is concentrated between 13℃ and 22℃, while the thermal radiation difference of intruding targets such as weasels is between 10℃ and 15℃. The thermal radiation difference caused by environmental disturbances is less than 10℃. Based on this, the thermal radiation gradient threshold is set to 13℃ to 22℃. According to actual measurements, the gait frequency of normal movement of chickens, ducks, and geese is concentrated between 10Hz and 30Hz, while the gait frequency of intruding weasels is between 35Hz and 60Hz. The frequency of equipment and environmental disturbances is low. Based on the gait frequency threshold of 10Hz, the gait frequency threshold is 10Hz to 30Hz. After preprocessing the historical sensing data, samples with thermal radiation gradients below 13℃ or above 22℃ and gait frequencies below 10Hz or above 30Hz are identified as invalid samples and removed. Among them, the thermal radiation gradients of invasive targets such as weasels are 10℃ to 15℃ and the gait frequencies are 35Hz to 60Hz, and they are all included in the invalid sample removal range. Finally, poultry candidate samples whose thermal radiation gradients and gait frequencies meet the threshold requirements are retained.

[0111] By preprocessing and dual threshold screening, invalid samples corresponding to environmental interference and intrusion targets are effectively eliminated, and poultry candidate samples are purified. This avoids the interference of invalid data on subsequent clustering analysis, improves the efficiency and accuracy of clustering operations, and serves as the core data carrier for subsequent density clustering with vibration spectrum and thermal radiation gradient as weighted features. The purity of the samples directly determines the effectiveness of density clustering in separating different types of poultry samples.

[0112] The candidate poultry samples contain different species, and the vibration spectrum and thermal radiation gradient of different species of poultry have characteristic differences. The density clustering algorithm does not require a preset number of clusters. It automatically classifies the samples based on the sample feature density, and uses the vibration spectrum and thermal radiation gradient as weighted features to maximize the highlighting of the characteristic differences of different species of poultry, so as to achieve accurate separation of different species of poultry samples.

[0113] Weighted features refer to assigning corresponding weights to vibration spectrum and thermal radiation gradient based on feature discrimination. Vibration spectrum has higher discrimination for poultry species and has a higher weight than thermal radiation gradient. Density clustering refers to a clustering algorithm based on sample feature density, which automatically divides samples with similar features into the same category. Different types of poultry samples refer to the sample sets corresponding to chickens, ducks, geese, etc., obtained after clustering.

[0114] For example, a weight of 0.6 is assigned to the vibration spectrum and a weight of 0.4 is assigned to the thermal radiation gradient. After normalizing the poultry candidate samples, a density clustering algorithm is used to perform clustering. The algorithm's neighborhood radius is set to 0.4 and the minimum number of core points is 4. The algorithm automatically classifies samples with vibration spectra of 10Hz to 20Hz and thermal radiation gradients of 18℃ to 22℃ as chicken samples, samples with vibration spectra of 15Hz to 25Hz and thermal radiation gradients of 16℃ to 20℃ as duck samples, and samples with vibration spectra of 20Hz to 30Hz and thermal radiation gradients of 13℃ to 18℃ as goose samples, thus completing the separation of different types of poultry samples.

[0115] By combining weighted features with density clustering, unsupervised and accurate separation of different types of poultry samples can be achieved, eliminating residual noisy samples and improving the purity of various poultry samples. This provides a clear data foundation for subsequent feature extraction, serving as a prerequisite for extracting body shape parameters, spatial coordinates, and gait frequencies and performing similarity clustering. The accuracy of sample classification directly affects the effect of subsequent similarity clustering.

[0116] The characteristics of single-category poultry samples still need to be further refined. Body size parameters, spatial coordinates and gait frequency are key features for distinguishing different types of poultry. Similarity clustering can quantify feature similarity to complete fine classification. Combined with the spatial zoning constraints of the breeding house, the clustering results can be matched with the actual breeding area, and finally the activity area, body size characteristics and gait frequency characteristics of different types of poultry can be obtained, providing a core basis for constructing a spatiotemporal behavior baseline.

[0117] Body shape parameters refer to the morphological parameters of poultry body length, width, and height extracted from point cloud features; spatial coordinates refer to the location coordinates of the poultry sample within the breeding house; gait frequency refers to the poultry movement frequency features extracted from the vibration spectrum; similarity clustering refers to clustering algorithms based on feature similarity; spatial partitioning constraints refer to the coordinate range of the breeding area divided by poultry species within the breeding house.

[0118] For example, the body shape parameters extracted from chicken samples are: body length 0.3m to 0.4m, width 0.2m to 0.3m, spatial coordinates concentrated in the 10m to 20m range, and gait frequency 10Hz to 20Hz; the body shape parameters extracted from duck samples are: body length 0.4m to 0.5m, width 0.3m to 0.4m, spatial coordinates concentrated in the 25m to 35m range, and gait frequency 15Hz to 25Hz; the body shape parameters extracted from goose samples are: body length... The spatial coordinates are concentrated in the 40m to 50m region, with a width of 0.6m to 0.7m and a gait frequency of 20Hz to 30Hz. After performing similarity clustering on the above features, combined with the spatial zoning constraints of the breeding house, the 10m to 20m region is divided into the chicken activity area, the 25m to 35m region into the duck activity area, and the 40m to 50m region into the goose activity area, so as to obtain the body shape characteristics and gait frequency characteristics of various poultry at the same time.

[0119] By combining feature extraction, similarity clustering, and spatial partitioning constraints, the characteristics of different types of poultry can be accurately extracted, clarifying the activity range, morphology, and movement characteristics of various types of poultry. This ensures that the feature results are consistent with the actual feeding layout of the poultry house, and serves as the basic data for subsequent fitting of thermal radiation gradients and construction of dynamic features. The accuracy of the features directly determines the effectiveness of the dynamic features.

[0120] The thermal radiation gradient of poultry fluctuates with changes in ambient temperature. Fixed thermal radiation features cannot adapt to different ambient temperature scenarios. By fitting gradient curves based on body surface thermal radiation data under multiple ambient temperatures, dynamic features that adapt to ambient temperature are constructed, enabling subsequent identification modules to accurately identify poultry under different ambient temperatures and improving the system's environmental adaptability.

[0121] Different ambient temperatures refer to the common temperature ranges in poultry houses, including low temperature, normal temperature and high temperature scenarios; body surface thermal radiation data refers to the thermal radiation values ​​of poultry body surfaces under different ambient temperatures; fitted thermal radiation gradient refers to the correspondence between ambient temperature and thermal radiation gradient obtained by fitting multiple sets of data; dynamic characteristics adapted to ambient temperature refer to the poultry thermal radiation characteristics that automatically adjust with changes in ambient temperature.

[0122] For example, three typical ambient temperatures of 10℃, 20℃, and 30℃ were selected to obtain surface thermal radiation data for chickens, ducks, and geese. At 10℃, the thermal radiation gradient for chickens was 20℃ to 22℃, for ducks it was 18℃ to 20℃, and for geese it was 16℃ to 18℃. At 20℃, the gradient for chickens was 18℃ to 20℃, for ducks it was 16℃ to 18℃, and for geese it was 14℃ to 16℃. At 30℃, the gradient for chickens was 16℃ to 18℃, for ducks it was 14℃ to 16℃, and for geese it was 13℃ to 15℃. Based on the above data, thermal radiation gradient curves were fitted to construct dynamic features for chickens, ducks, and geese that were adapted to the ambient temperature, allowing the features to be adjusted in real time with the ambient temperature.

[0123] This allows for the construction of dynamic poultry features that adapt to ambient temperature, addressing the issue of thermal radiation feature recognition failure caused by changes in ambient temperature. It also improves the system's recognition stability under different farming environments. The completed dynamic features will be incorporated into the construction elements of the subsequent spatiotemporal behavior baseline, providing a dynamic comparison standard for hierarchical matching of the recognition module.

[0124] Specifically, similarity clustering is performed for:

[0125] Calculate the similarity of body shape parameters and spatial coordinate distance of corresponding point cloud features in poultry samples, and cluster the poultry samples using Euclidean distance;

[0126] Based on the spatial zoning constraints of the poultry house, the activity areas and body size characteristics of different types of poultry are matched;

[0127] Gait frequencies corresponding to poultry samples are extracted, and the poultry samples are clustered using a density clustering algorithm. The consistency of gait frequencies in different clustering results is compared to obtain the gait frequency characteristics of different types of poultry.

[0128] Different types of poultry have fixed differences in body shape parameters and spatial distribution. Body shape parameter similarity can quantify the degree of similarity in poultry morphology, and spatial coordinate distance can reflect the distance of poultry activity locations. Euclidean distance algorithm can accurately quantify the comprehensive feature differences between samples. By performing clustering through this algorithm, poultry samples with similar morphology and location can be divided into the same category, providing a clustering basis for subsequent matching of activity areas and body shape features.

[0129] Body shape parameter similarity refers to the degree of similarity of body shape parameters among different poultry samples calculated by cosine similarity; spatial coordinate distance refers to the straight-line distance between the spatial coordinates of different poultry samples; Euclidean distance refers to the calculation method that quantifies the comprehensive difference between samples after weighted fusion of body shape parameter similarity and spatial coordinate distance; similarity clustering refers to the clustering method that divides samples with small comprehensive differences into one class based on Euclidean distance.

[0130] For example, two groups of chicken samples and two groups of duck samples are selected. The cosine similarity of the body shape parameters of the first group of chicken samples and the second group of chicken samples is 0.96, and the Euclidean distance of the spatial coordinates is 1.2m. Since the overall Euclidean distance is small, they are divided into the same cluster. The cosine similarity of the body shape parameters of the chicken samples and the duck samples is 0.71, and the Euclidean distance of the spatial coordinates is 15m. Since the overall Euclidean distance is large, they are divided into different clusters. Finally, the clusters of chickens, ducks and geese are obtained by Euclidean distance clustering.

[0131] By quantifying the dual features of body shape parameters and spatial coordinates, and combining them with the Euclidean distance algorithm, poultry samples can be accurately clustered. The clustering results fit the morphology and spatial distribution characteristics of poultry. The poultry sample clusters obtained by clustering are the core objects for subsequent matching of activity areas and body shape features by combining spatial partition constraints. The clustering accuracy directly determines the accuracy of feature matching.

[0132] Similarity clustering can only obtain the feature classification results of the samples and cannot directly associate poultry species with the actual breeding area. The spatial zoning constraint of the breeding house is the breeding area divided according to the poultry species. Combining this constraint, the clusters are accurately matched with the poultry species and activity areas, and the mean body size parameters of the clusters are extracted to obtain the standard body size characteristics of different types of poultry.

[0133] Spatial partition constraints refer to the spatial coordinate range of the breeding houses corresponding to different types of poultry, which is preset by the system; matching refers to comparing the mean spatial coordinates of the clusters with the spatial partition coordinate ranges to determine the poultry species corresponding to the clusters; activity area refers to the poultry breeding space range determined after matching; body shape characteristics refer to the mean characteristics of body shape parameters within the clusters.

[0134] For example, the spatial zoning constraints of the breeding shed are 10m to 20m for chickens, 25m to 35m for ducks, and 40m to 50m for geese. The average spatial coordinate of the first cluster is 15m, which falls within the chicken zoning range and is matched as a chicken sample. The average body shape parameters of this cluster are extracted as body length of 0.35m and width of 0.25m, which are used as the body shape characteristics of chickens. The activity area is 10m to 20m. The average spatial coordinate of the second cluster is 30m, which is matched as a duck sample. The average spatial coordinate of the third cluster is 45m, which is matched as a goose sample. The corresponding body shape characteristics and activity areas are obtained simultaneously.

[0135] This enables precise matching between clustering results and actual farming scenarios, clearly defining the activity areas and standard body shape characteristics of different types of poultry. The feature results have practical application value, aligning with the layout requirements of large-scale farming. The matched activity areas and body shape characteristics provide a well-defined sample basis for subsequent gait frequency extraction and cluster analysis, ensuring the targeted extraction of gait frequency features.

[0136] Different types of poultry exhibit fixed frequency differences in their gait patterns. Density clustering algorithms can automatically classify gait patterns based on the density distribution of these frequencies. By comparing the consistency of gait frequencies among different clustering results, the optimal clustering result can be selected, ultimately yielding the standard gait frequency characteristics of different types of poultry and improving the core behavioral characteristic system of poultry.

[0137] Gait frequency refers to the core frequency corresponding to the vibration spectrum of poultry during movement; density clustering algorithm refers to an unsupervised clustering algorithm based on gait frequency density distribution; gait frequency consistency refers to the similarity of gait frequencies of samples within the same cluster; gait frequency feature refers to the range of gait frequencies corresponding to the optimal clustering result.

[0138] For example, the gait frequency of chicken samples is extracted as 10Hz to 20Hz, that of duck samples as 15Hz to 25Hz, and that of goose samples as 20Hz to 30Hz. The density clustering algorithm is used to cluster the gait frequencies, resulting in multiple candidate clustering results. The consistency of the gait frequencies of each group is compared, and the clustering result with a consistency of more than 95% is the optimal result. Finally, the gait frequency features of chickens are extracted as 10Hz to 20Hz, that of ducks as 15Hz to 25Hz, and that of geese as 20Hz to 30Hz.

[0139] By using density clustering and consistency comparison, standard gait frequency characteristics of different types of poultry are obtained. These characteristics are highly stable and discriminative, and a complete three-dimensional key feature of poultry morphology, space and movement is constructed. The obtained gait frequency characteristics of different types of poultry, together with activity area, body size characteristics and dynamic characteristics, will serve as behavioral benchmark features, providing complete feature support for constructing a spatiotemporal behavioral baseline.

[0140] Specifically, such as Figure 2 As shown, the spatiotemporal behavioral baseline is constructed for:

[0141] The activity area, body size characteristics, and gait frequency characteristics of different types of poultry were set as the behavioral benchmark characteristics of poultry.

[0142] By binding the gait frequency characteristics of the same type of poultry with the corresponding activity area and body size characteristics, spatiotemporal association pairs are formed for each type of poultry.

[0143] Based on historically acquired perception data, dynamic threshold ranges are set for the corresponding behavioral baseline features and dynamic features of each type of poultry in terms of spatiotemporal correlation.

[0144] For each type of poultry, a sub-baseline is established, which includes corresponding behavioral baseline features, dynamic features, dynamic threshold ranges, and spatiotemporal correlation rules. All sub-baselines are integrated to form a spatiotemporal behavioral baseline.

[0145] The previous cluster analysis and similarity clustering have completed the extraction of key features of different types of poultry. However, these features are only raw statistical results and have not been clearly defined as the standard basis for subsequent identification and comparison. They cannot be directly used to distinguish legitimate poultry from invasive targets such as weasels. Setting the extracted features as behavioral benchmark features can establish a standardized reference system for legitimate targets, providing a core basis for subsequent spatiotemporal correlation, threshold setting and baseline construction.

[0146] Behavioral baseline characteristics refer to standardized reference characteristics used to determine whether a target is legitimate poultry. They consist of three categories: activity area, body shape characteristics, and gait frequency characteristics. The activity area refers to the range of normal activity space for different types of poultry in the breeding house. Body shape characteristics refer to the average morphological parameters such as body length and width corresponding to the dot cloud of different types of poultry. Gait frequency characteristics refer to the range of core frequencies of the vibration spectrum corresponding to the normal movement of different types of poultry.

[0147] For example, the activity area of ​​chickens (10m to 20m), body shape characteristics (body length 0.3m to 0.4m and width 0.2m to 0.3m), and gait frequency characteristics (10Hz to 20Hz) obtained from clustering are uniformly set as the behavioral baseline characteristics of chickens; the activity area of ​​ducks (25m to 35m), body shape characteristics (body length 0.4m to 0.5m and width 0.3m to 0.4m), and gait frequency characteristics (15Hz to 25Hz) are set as the behavioral baseline characteristics of ducks; and the activity area of ​​geese (40m to 50m), body shape characteristics (body length 0.6m to 0.7m and width 0.4m to 0.5m), and gait frequency characteristics (20Hz to 30Hz) are set as the behavioral baseline characteristics of geese.

[0148] This process standardizes and benchmarks the characteristics of different types of poultry, clarifies the core reference standards for legitimate poultry, distinguishes them from invasive targets such as weasels, and provides a stable characteristic basis for the formation of subsequent spatiotemporal association pairs. The established behavioral benchmark characteristics are the only material for binding the characteristics of the same type of poultry to form spatiotemporal association pairs. The accuracy of the benchmark characteristics directly determines the rationality of the spatiotemporal association pairs.

[0149] Individual behavioral baseline features are independent of each other and cannot reflect the spatiotemporal correlation of poultry behavior. That is, the gait frequency of poultry must correspond to and match its own activity area and body shape characteristics. If there is a situation where the gait frequency matches a certain type of poultry but the activity area and body shape characteristics do not match, the legitimacy of the target cannot be accurately determined. Binding the three types of features of the same type of poultry to form a spatiotemporal correlation pair can establish a fixed correspondence between features and strengthen the uniqueness and relevance of the features of legitimate targets.

[0150] Spatiotemporal association refers to uniquely binding the gait frequency characteristics, activity area, and body shape characteristics of the same type of poultry, forming a combination of movement characteristics, spatial characteristics, and morphological characteristics, ensuring that the three types of characteristics belong to the same type of poultry; binding means assigning a unified identifier to the three types of characteristics of each type of poultry, establishing an inseparable correspondence.

[0151] The gait frequency characteristics of chickens (10Hz to 20Hz) were bound to their corresponding activity areas (10m to 20m) and body shape characteristics (body length 0.3m to 0.4m and width 0.2m to 0.3m) to form spatiotemporal association pairs. Similarly, the gait frequency characteristics of ducks (15Hz to 25Hz) were bound to their corresponding activity areas (25m to 35m) and body shape characteristics (body length 0.4m to 0.5m and width 0.3m to 0.4m) to form spatiotemporal association pairs. Likewise, the gait frequency characteristics of geese (20Hz to 30Hz) were bound to their corresponding activity areas (40m to 50m) and body shape characteristics (body length 0.6m to 0.7m and width 0.4m to 0.5m) to form spatiotemporal association pairs.

[0152] This establishes a fixed spatiotemporal association of three types of behavioral baseline characteristics of poultry of the same kind, avoiding feature confusion and mismatch, ensuring that the feature combination of legitimate targets is unique, effectively distinguishing them from the scattered abnormal features of intrusive targets such as weasels, and improving the accuracy of subsequent identification. The spatiotemporal association pairs formed for each type of poultry are the direct objects for setting dynamic threshold intervals based on historical perception data. The completeness of the association pairs determines the pertinence of the threshold interval setting.

[0153] The behavioral baseline characteristics and dynamic characteristics of poultry adapted to environmental temperature are not fixed and will fluctuate normally with poultry growth, changes in environmental temperature, and movement status. Fixed thresholds will judge normal fluctuations as abnormal, leading to system misjudgment. Setting dynamic threshold ranges based on historical perception data can cover the normal fluctuation range of poultry characteristics, while excluding the characteristic range of intrusive targets such as weasels, ensuring the compatibility and accuracy of identification.

[0154] The dynamic threshold range refers to the numerical range that allows for normal fluctuations in poultry characteristics, obtained based on historical sensing data statistics, and can be adaptively adjusted according to the scenario; dynamic characteristics refer to the thermal radiation gradient characteristics of different types of poultry adapted to environmental temperatures as described above; historically acquired sensing data refers to the effective sensing data of legal poultry collected by the system over a long period of time.

[0155] For example, based on historical perception data, dynamic threshold ranges are set for the spatiotemporal correlation of chickens: activity area 9m to 21m, body shape characteristics: body length 0.25m to 0.45m and width 0.15m to 0.35m, gait frequency characteristics 8Hz to 22Hz, and dynamic characteristic: thermal radiation gradient 13℃ to 22℃; dynamic threshold ranges are set for ducks: activity area 24m to 36m, body shape characteristics: body length 0.35m to 0.55m and width 0.25m to 0.35m. The following dynamic threshold ranges were set for geese: 0.45m, gait frequency characteristics 13Hz to 27Hz, dynamic characteristic thermal radiation gradient 13℃ to 20℃; and an activity area of ​​39m to 51m, body size characteristics of 0.55m to 0.75m body length and 0.35m to 0.55m width, gait frequency characteristics of 18Hz to 32Hz, and dynamic characteristic thermal radiation gradient of 13℃ to 18℃. All threshold ranges covered the normal fluctuations of poultry characteristics and excluded the characteristic range of weasels.

[0156] This allows for setting a reasonable dynamic fluctuation range for the spatiotemporal correlation and dynamic characteristics of poultry. It allows for fluctuations in the normal characteristics of legitimate poultry while accurately filtering out abnormal characteristics of intruding targets, reducing the system's misjudgment rate, and improving environmental adaptability. The set dynamic threshold range, together with behavioral baseline characteristics and dynamic characteristics, constitutes the core elements of the sub-baseline and is a necessary component for establishing sub-baselines for each type of poultry.

[0157] Behavioral baseline features, dynamic features, dynamic threshold ranges, and spatiotemporal association rules are still scattered elements and cannot be directly used for hierarchical matching of the identification module. To establish an independent sub-baseline for each type of poultry, the four types of elements are integrated into a standardized identification unit, and then all sub-baselines are integrated to form a complete spatiotemporal behavioral baseline, providing a legitimate target reference system for the identification module.

[0158] A sub-baseline refers to an independent identification baseline established for a single species of poultry, which includes four types of elements: behavioral baseline features, dynamic features, dynamic threshold ranges, and spatiotemporal association rules. Spatiotemporal association rules refer to the judgment rules defined in the sub-baseline, that is, the same target must simultaneously meet the threshold requirements of the three types of features in the spatiotemporal association pair and match the corresponding dynamic features. The spatiotemporal behavioral baseline refers to a complete legal target behavior reference system formed by integrating all poultry sub-baselines.

[0159] For example, a sub-baseline is established for chickens, which includes the chicken's behavioral baseline characteristics, dynamic characteristics adapted to environmental temperature, corresponding dynamic threshold ranges, and spatiotemporal association rules that require the target to be simultaneously within the threshold range of the chicken's activity area, with body shape and gait frequency characteristics meeting the thresholds, and matching the chicken's dynamic thermal radiation gradient. Similarly, dedicated sub-baselines are established for ducks and geese. The three sub-baselines for chickens, ducks, and geese are integrated according to the spatial zoning of the breeding house to form a spatiotemporal behavioral baseline covering all legal poultry, which serves as the sole reference standard for hierarchical matching of the identification module.

[0160] This integrates scattered features and rules into a systematic and standardized spatiotemporal behavior baseline, forming a complete legitimate target identification system. This directly supports the hierarchical matching process of the identification module, enabling accurate differentiation between legitimate and intrusive targets. The completed spatiotemporal behavior baseline is the core reference for the identification module to extract multidimensional features of moving targets and perform hierarchical matching. The completeness and accuracy of the baseline directly determine the effectiveness of subsequent target identification and intrusion determination.

[0161] The identification module is used to extract moving targets from the current perception data and obtain the multi-dimensional features of the moving targets. The multi-dimensional features are then matched hierarchically with the spatiotemporal behavior baseline. If the match is successful, the moving target is determined to be a legitimate target; if the match fails, the moving target is determined to be an intrusion target.

[0162] Furthermore, the multidimensional features of the moving target are obtained for:

[0163] The current sensing data is preprocessed, and the core area of ​​the moving target is located based on the feature of the point cloud, with the moving target as the dynamic background. The thermal radiation gradient and vibration spectrum corresponding to the core area are then locked.

[0164] Gait frequency features are extracted from the locked vibration spectrum, body shape parameters are extracted from the point cloud features, and gradient values ​​adapted to the current ambient temperature are extracted from the thermal radiation gradient as core features.

[0165] Extract the state of the moving target point cloud and the real-time spatial coordinates of the moving target from the current sensing data as auxiliary features;

[0166] The core features and auxiliary features are verified to determine whether they belong to the same moving target. Based on the gait frequency features, the verified core features and auxiliary features are correlated and normalized to form a multidimensional feature of the moving target.

[0167] The real-time acquired current sensing data includes static environmental data such as the walls of the breeding house, feeding equipment and poultry house supports. At the same time, there is a feature intersection problem caused by the coexistence of multiple targets. Unprocessed raw data will lead to feature extraction deviation and target confusion. Processing with moving targets as dynamic background can eliminate static redundant information, focus on effective target areas, and ensure the targeting of subsequent feature extraction.

[0168] The current sensing data refers to the fused data output in real time by the sensing module, which includes point cloud features, thermal radiation gradient, and vibration spectrum. Preprocessing uses a Gaussian filtering algorithm to remove random noise from the data and regularizes data with missing local features. Using the moving target as a dynamic background means setting the fixed facilities in the breeding house as a static background and filtering them out, retaining only the point cloud data, thermal imaging data, and vibration data corresponding to the moving target. The core region refers to the local area defined by the geometric center of the moving target's point cloud, which is the spatial range where the target features are most concentrated. Locking the corresponding features means uniquely matching the spatial location of the core region with the thermal radiation gradient and vibration spectrum through spatial coordinate mapping.

[0169] For example, the point cloud acquisition range of the lidar in the breeding house is 0m to 10m, 0m to 5m, and 0m to 2.5m, and the temperature measurement range of the infrared thermal imager is 0℃ to 50℃. After preprocessing, the static facility point cloud is filtered out, and the core area is located with the moving target as the dynamic background. The spatial coordinates of the core area of ​​the chicken are 15m, 2.5m, and 0.3m, and the spatial coordinates of the core area of ​​the weasel are 5m, 2.5m, and 0.2m. Based on the coordinates of the core area, the corresponding thermal radiation gradient and vibration spectrum data are locked.

[0170] This process eliminates redundant information from the static environment, removes interference from the cross-interference of multiple target features, and accurately locates the core feature region of the moving target, providing a data source for subsequent feature extraction. The core region and the locked feature data are then used as the processing objects for extracting core features. The accuracy of the core region location determines the accuracy of the core feature extraction.

[0171] Gait frequency, body shape parameters, and thermal radiation gradient values ​​of the suitable ambient temperature are highly discriminative features that distinguish poultry from invasive targets. These features can uniquely identify the target type from the dimensions of movement characteristics, morphological characteristics, and thermal characteristics, and are the core basis for subsequent hierarchical matching. They need to be extracted from the locked feature data.

[0172] Gait frequency characteristics refer to the core frequency values ​​of the moving target extracted from the vibration spectrum, which are used to characterize the rhythm characteristics of the movement; body shape parameters refer to the morphological parameters such as body length and body width extracted from the point cloud features, which are used to characterize the target's external dimensions; gradient values ​​adapted to the current ambient temperature refer to the corresponding gradient values ​​retrieved from the thermal radiation gradient according to the current ambient temperature, which match the dynamic thermal radiation characteristics of poultry.

[0173] For example, the gait frequency characteristics of chickens range from 10Hz to 20Hz, with body length parameters of 0.3m to 0.4m and body width parameters of 0.2m to 0.3m, and a thermal radiation gradient value suitable for an ambient temperature of 20℃ of 18℃ to 20℃; the gait frequency characteristics of ducks range from 15Hz to 25Hz, with body length parameters of 0.4m to 0.5m and body width parameters of 0.3m to 0.4m; and the gait frequency characteristics of weasels range from 35Hz to 60Hz, with body length parameters of 0.25m to 0.45m and body width parameters of 0.12m to 0.2m. The above extraction results are all used as the core features of the corresponding targets.

[0174] This enables the targeted extraction of highly discriminative core features, constructs the core feature dimension for target recognition, highlights the feature differences between poultry and invasive targets, provides a core reference for subsequent auxiliary feature extraction and feature verification, and serves as the benchmark feature for feature verification and normalization processing.

[0175] Relying solely on core features cannot cover all recognition scenarios. In some scenarios, core features overlap. The cloud state of moving target points and real-time spatial coordinates are spatiotemporal auxiliary features that can supplement the judgment dimension of core features, improve the reliability of recognition results, and reduce the probability of misjudgment.

[0176] The state of the moving target point cloud refers to the aggregation, dispersion, and other morphological manifestations of the point cloud, which is used to reflect the movement status of the target; the real-time spatial coordinates refer to the position values ​​of the core area of ​​the moving target in the spatial coordinate system of the breeding house at the current moment, which are used to characterize the spatial location of the target.

[0177] For example, the point cloud of chickens is mainly in a clustered state, with real-time spatial coordinates stable in the range of 10m to 20m; the point cloud of weasels is in a discrete state, with real-time spatial coordinates mostly in the inlet and outlet passage area from 0m to 5m; the point cloud of ducks is in a loosely clustered state, with real-time spatial coordinates in the range of 25m to 35m. The above extraction results are all used as auxiliary features for the corresponding targets.

[0178] This supplements the spatiotemporal auxiliary features, improves the target feature system, makes up for the lack of differentiation of core features in special scenarios, enhances the overall feature recognition adaptability, and participates in feature attribution verification together with core features. The effectiveness of auxiliary features directly affects the reliability of verification results.

[0179] In scenarios where multiple targets coexist, core features and auxiliary features may belong to different targets. This confusion in feature attribution can directly lead to recognition errors. Furthermore, the dimensions of core features and auxiliary features are not consistent, making them unsuitable for hierarchical matching. Verification and association normalization can eliminate these problems and form standardized multidimensional features.

[0180] Feature attribution verification adopts a spatiotemporal consistency verification method to check whether the timestamps and spatial coordinates of core features and auxiliary features are consistent. If they are consistent, they are determined to belong to the same target; if they are inconsistent, they are determined to be feature confusion and are removed. Association normalization processing refers to using gait frequency features as a benchmark to transform core features and auxiliary features of different dimensions into a unified numerical range, eliminating the difference in dimensions and maintaining the relative difference relationship between features. Multidimensional features refer to the standardized feature set formed by integrating core features and auxiliary features after verification and normalization.

[0181] For example, if the core and auxiliary features of a chicken have the same timestamp and spatial coordinates of 15m, 2.5m, and 0.3m, they are verified to belong to the same target. Using the gait frequency feature of 10Hz to 20Hz as a benchmark, the body shape parameters, thermal radiation gradient values, point cloud state, and real-time spatial coordinates are converted to a unified range and integrated to form the multidimensional features of the chicken. If the spatial coordinates of the core features are 5m, 2.5m, and 0.2m, and the spatial coordinates of the auxiliary features are 15m, 2.5m, and 0.3m, and the spatiotemporal coordinates are inconsistent, then the features are judged to be confused, and this set of features is not used.

[0182] This eliminates feature confusion, ensures the uniqueness of feature sources, and unifies feature dimensions through normalization, forming standardized multidimensional features that can be used for hierarchical matching. This provides unique feature input data for the hierarchical matching step, and the degree of standardization and accuracy of the multidimensional features directly determines the accuracy of hierarchical matching.

[0183] Specifically, hierarchical matching is used for:

[0184] The core features of the multidimensional features are compared with the behavioral benchmark features of the corresponding sub-baseline dimension by dimension to determine whether the core features are within the dynamic threshold range.

[0185] If the core feature is not within the dynamic threshold range, the current sub-baseline is directly determined to be a failure to match, and the matching of other types of poultry sub-baselines is switched.

[0186] If the core feature is within the dynamic threshold range, the auxiliary features of the multidimensional feature are compared with the behavioral baseline features, spatiotemporal association rules, and the dynamic features of the corresponding sub-baseline.

[0187] If the auxiliary feature comparison is successful, then verify whether the multidimensional features involved in the matching are valid features of the same moving target;

[0188] If the verification passes, the match is considered successful and the corresponding moving target is determined to be a legitimate target. If the verification fails or the auxiliary feature comparison is unqualified, the corresponding sub-baseline is considered to have failed to match. If all sub-baselines fail to match, the moving target is considered to be an intrusion target.

[0189] The core features are the basic identification dimensions for distinguishing legitimate aquaculture targets. They are the core behavioral characteristics that determine whether a moving target conforms to the legitimate target. Initial screening needs to be completed by comparing each dimension to exclude abnormal objects that deviate significantly from the characteristics of legitimate targets, thereby reducing the redundant computation of subsequent identification processes. The hierarchical identification logic of first identifying the core dimensions and then the auxiliary dimensions should be followed.

[0190] Multidimensional features are a set of standardized features extracted and normalized. Core features refer to three basic identification features: gait frequency, body shape parameters, and thermal radiation gradient values. Sub-baselines are independent behavioral baselines established for each type of legal poultry in the breeding house. Behavioral baseline features are the feature baseline parameters corresponding to the normal behavior of legal poultry pre-set in the sub-baselines. The dynamic threshold range is the allowable fluctuation range of core features adjusted according to changes in the breeding environment and the poultry growth cycle. The core features in the multidimensional features acquired in real time are compared with the preset behavioral baseline features in the currently invoked sub-baselines, one by one, to check whether each core feature falls within the corresponding dynamic threshold range.

[0191] For example, in the sub-baseline corresponding to chickens in the breeding house, the behavioral baseline features include gait frequency of 10Hz to 20Hz, body length of 0.3m to 0.4m, and thermal radiation gradient value adapted to the corresponding environmental gradient. The dynamic threshold range is the fluctuation range of the behavioral baseline features within 15%. The core features extracted in real time (gait frequency of 12Hz, body length of 0.35m, and thermal radiation gradient value adaptation) are checked against the behavioral baseline features of the chicken sub-baseline dimension by dimension to confirm that each core feature falls within the corresponding dynamic threshold range.

[0192] This allows for the use of core features as the initial screening criterion, quickly eliminating moving targets that clearly do not conform to the characteristics of legitimate targets, reducing invalid processes in the system's identification, and ensuring the accuracy and efficiency of the initial screening process.

[0193] Each sub-baseline corresponds to the behavioral characteristics of a single legal poultry. When the core feature is not within the dynamic threshold range of the corresponding sub-baseline, it indicates that the current moving target does not belong to the legal poultry corresponding to that sub-baseline. It is necessary to switch to the sub-baseline corresponding to other types of legal poultry in the breeding house for further comparison, so as to avoid directly judging the target as abnormal due to the failure of a single sub-baseline match, and to cover the identification range of all legal poultry in the breeding house.

[0194] Sub-baseline matching failure means that after comparing the current core feature with the behavioral benchmark feature of the corresponding sub-baseline dimension by dimension, at least one core feature exceeds the dynamic threshold range, and the current sub-baseline is automatically determined to be unable to complete the matching of the current moving target. Switching to match other types of poultry sub-baselines is to call the sub-baselines corresponding to other legal poultry such as ducks and geese in the breeding house in the order of preset legal poultry types, and re-execute the dimension-by-dimensional comparison process of core features and behavioral benchmark features.

[0195] For example, after comparing the core features of the current moving target with the chicken baseline, if the gait frequency of 25Hz exceeds the corresponding dynamic threshold range, it is determined that the chicken baseline matching has failed. Then, it automatically switches to the sub-baseline corresponding to the duck in the breeding house and restarts the dimensional comparison process of the core features.

[0196] By using a multi-sub-baseline polling comparison method, the core feature range of all legal poultry in the breeding house can be fully covered, avoiding misjudgment of legal targets due to the limitation of a single sub-baseline, improving the comprehensiveness of system identification. The execution result directly affects the benchmark object of the subsequent comparison process. The switched sub-baseline serves as the new core feature comparison benchmark, and the initial screening process of core features is re-executed until all legal poultry sub-baselines have been polled or matched.

[0197] Relying solely on core features is insufficient for comprehensive identification of moving targets. Auxiliary features include spatiotemporal correlation, point cloud state, and other supplementary identification dimensions. These features must be combined with the pre-set behavioral baseline features, spatiotemporal correlation rules, and dynamic features of the sub-baseline for comprehensive comparison to improve the identification dimension system, eliminate the limitations of identification based on a single core feature, and ensure the accuracy of the identification results.

[0198] Auxiliary features refer to two types of identification features: the extracted cloud state of moving target points and their real-time spatial coordinates. Behavioral baseline features refer to the preset legal poultry auxiliary behavioral baseline parameters in the corresponding sub-baseline. Spatiotemporal association rules refer to the preset corresponding constraint rules for the activity positions and behavioral features of legal poultry in the corresponding sub-baseline. Dynamic features refer to the thermal radiation gradient features of different types of poultry adapted to environmental temperatures. The real-time acquired auxiliary features are compared one by one with the preset behavioral baseline features, spatiotemporal association rules, and dynamic features in the current sub-baseline to verify whether the auxiliary features meet all preset constraint conditions.

[0199] For example, if the core features meet the dynamic threshold range requirements of the chicken baseline, the auxiliary features extracted in real time (point cloud cluster aggregation state, real-time spatial coordinates 15m to 20m) are compared with the preset behavioral benchmark features, spatiotemporal association rules (normal chicken activity spatial coordinates 10m to 20m, point cloud cluster aggregation state is normal) and dynamic features in the chicken baseline to confirm that the auxiliary features meet all constraints.

[0200] By comparing auxiliary features in multiple dimensions, the identification dimensions of core features are supplemented, the loopholes in the identification of a single core feature are eliminated, the integrity and reliability of the system identification are improved, and the conditions for starting the subsequent feature validity verification process are determined.

[0201] Passing the auxiliary feature comparison only indicates that the feature parameters conform to the preset rules. It cannot guarantee that the core features and auxiliary features involved in the comparison belong to the same moving target. Feature confusion is likely to occur in the scenario of multiple targets coexisting in the breeding house. It is necessary to eliminate the situation of feature confusion through validity verification to ensure that the feature source of the identification basis is unique and valid.

[0202] Valid features refer to multi-dimensional features corresponding to the same moving target with consistent timestamp information, matching spatial coordinate information, and unique feature sources. The verification process refers to the system retrieving the timestamp information and spatial coordinate information of the multi-dimensional features, checking whether the timestamps of the core features and auxiliary features are completely consistent and whether the spatial coordinates are completely matched, and confirming that all multi-dimensional features participating in the comparison belong to the same moving target, without feature confusion or cross-target splicing.

[0203] For example, after the auxiliary features pass the comparison, the timestamp information (both are collected at the same time) and spatial coordinate information (both are in the range of 15m to 20m) of the core features and auxiliary features are retrieved to complete the attribution verification of all features and confirm that there is no feature confusion.

[0204] By using the feature validity verification process, the misidentification problem caused by the confusion of multiple target features is eliminated, ensuring that the feature source on which the system identifies is unique and valid, thereby improving the accuracy of the system's identification and determining the final matching result. If the verification passes, the corresponding sub-baseline is determined to be successfully matched; if the verification fails, the corresponding sub-baseline is determined to be unmatched.

[0205] After completing the comparison and verification of a single sub-baseline, all legal poultry sub-baselines need to be polled for confirmation. If all sub-baselines fail to match, it means that the moving target does not belong to any legal poultry in the breeding house and is judged as an intrusion target. This achieves the final accurate distinction between legal targets and intrusion targets, providing a reliable basis for judgment in the subsequent alarm process.

[0206] A successful match means that the feature validity check passes and the current moving target is determined to be a legitimate poultry target corresponding to the sub-baseline. A sub-baseline match failure means that the auxiliary feature comparison is unqualified or the feature validity check fails, and the process switches to the next legitimate poultry sub-baseline to continue the comparison process. All sub-baselines fail to match means that after polling all legitimate poultry sub-baselines in the poultry house, such as chickens, ducks, and geese, there are no matching results, and the current moving target is automatically determined to be an intrusion target.

[0207] For example, after polling all legal poultry sub-baselines of chickens, ducks, and geese in the breeding house, if the core features or auxiliary features do not meet the corresponding dynamic threshold range and comparison rules, the current moving target is determined to be an intrusion target.

[0208] By using full sub-baseline polling and final judgment process, the system can accurately distinguish between legitimate targets and intrusion targets, ensuring the final accuracy of system identification and providing reliable judgment support for subsequent intrusion alarm processes. This determines the execution of subsequent actions; if a target is determined to be legitimate, the system will not trigger an alarm process, while if it is determined to be an intrusion target, the intrusion alarm process will be triggered immediately.

[0209] The response module is used to execute a graded response when the moving target is determined to be an intrusion target or when an abnormal cloud of moving target points is detected. When the intrusion target persists or a preset destructive event is triggered, alarm information is sent and physical defense facilities are activated.

[0210] Specifically, the response module is used for:

[0211] When a moving target is determined to be an intrusion target or an anomaly is detected in the point cloud of a moving target, a graded response is immediately executed to carry out corresponding expulsion actions, and the existence status of the intrusion target and whether a preset destruction event is triggered are monitored in real time.

[0212] If the intrusion target persists or a preset sabotage event is triggered, a corresponding alarm message will be generated and sent out, and the linkage defense action of the physical defense facilities will be activated at the same time.

[0213] If the intrusion target disappears after being driven away and no preset destruction event is triggered, the driving action will automatically stop and no alarm information will be sent or physical defense facilities will be linked.

[0214] Point cloud anomalies are a critical abnormal state in aquaculture scenarios that can easily lead to target misjudgment and failure of area control. If graded response and expulsion actions are not initiated in a timely manner, it will further lead to risks such as poultry panic and failure of aquaculture area control. At the same time, it is necessary to monitor the existence status of intruding targets and whether preset sabotage events are triggered in real time during the expulsion action. Preset sabotage events need to be refined and graded according to the aquaculture scenario to achieve accurate matching of alarm and response.

[0215] Point cloud anomaly refers to an abnormal state in which the speed of change or movement of a moving target's point cloud exceeds a preset normal threshold. The preset normal threshold is determined based on statistical data of the normal behaviors of legitimate poultry such as chickens, ducks, and geese in the breeding shed, including feeding, walking, and running within a small area. It summarizes the speed parameters of a large number of normal samples obtained over a long period of time, and takes the upper limit of the speed of normal poultry activities as the judgment threshold. At the same time, it combines the characteristics of intruding targets such as weasels, which move faster, to effectively distinguish between normal poultry behavior and abnormal movement states. Tiered response refers to the tiered handling response initiated according to the level of target anomaly and the risk level of the breeding scenario. The driving action adopts sound and light-based driving action, including strong light driving action and animal husbandry voice broadcast driving action. Strong light driving action refers to driving away animals by emitting strong light signals through high-brightness lights deployed in the breeding area. Animal husbandry voice broadcast driving action refers to driving away animals by broadcasting animal husbandry voice signals through audio equipment deployed in the breeding area. Both types of driving action are in line with the breeding scenario and will not cause secondary harm to poultry.

[0216] The pre-set sabotage events are categorized into three levels based on the farming scenario. Level 1 is ordinary poultry panic, which refers to a panic state lasting only a few seconds, accompanied by abnormal noise, followed by a rapid return to quiet and no recurrence within a certain period. Level 2 is external intrusion but unsuccessful intrusion, which refers to the presence of abnormal frequencies at the farm entrance / exit, abnormal temperature detected by infrared sensors, abnormal point cloud detected by video, accompanied by abnormal noise, but the intruding target fails to enter the core farming area. Level 3 is external intrusion and successful intrusion, which refers to the abnormally rapid gathering, dispersal, and movement of poultry, accompanied by abnormal noise, abnormal frequencies at the farm entrance / exit, abnormal temperature detected by infrared sensors, and abnormal point cloud detected by video, and the intruding target has successfully entered the core farming area. When a moving target is determined to be an intruding target, or when an abnormal state of a point cloud is detected where the movement speed of the moving target's point cloud is too fast or exceeds the pre-set normal threshold, a graded response is immediately initiated, implementing sound and light-based deterrent actions, while simultaneously monitoring the presence of the intruding target in real time, and whether the above three levels of pre-set sabotage events have been triggered.

[0217] For example, based on actual measurement results, the normal change rate threshold for point cloud clusters is set to 0.15 m / s, and the movement speed threshold for moving targets is set to 0.8 m / s. Both speed parameters for normal activities of chickens, ducks, and geese are consistently lower than the above values. The breeding area has preset thresholds for the normal change rate and movement speed of point cloud clusters, and the point cloud cluster status of moving targets is monitored in real time. When the point cloud cluster change rate or movement speed of a moving target is detected to be too fast or exceeds the corresponding preset threshold, it is determined to be an abnormal point cloud cluster, or the moving target is directly determined to be an intruding target. The graded response is immediately initiated, and strong light irradiation is turned on to carry out strong light irradiation to drive away the target. At the same time, the voice signal of the breeder is broadcast in a loop through the audio equipment to carry out the breeder's voice broadcast to drive away the target. Simultaneously, it is monitored in real time whether the intruding target continues to exist in the breeding area, and whether it triggers three preset damage events: ordinary poultry panic, unsuccessful external intrusion, and successful external intrusion.

[0218] This enables precise identification of two triggering conditions: intrusion targets and abnormal point cloud clusters. It allows for timely initiation of sound and light-based deterrence actions tailored to the aquaculture scenario, with a gentle deterrence method that causes no secondary harm. Simultaneously, it precisely refines three levels of preset destructive events, enabling accurate differentiation of risk states in aquaculture scenarios and providing a precise basis for subsequent graded handling.

[0219] The pre-set sabotage events have been refined into three levels: ordinary poultry panic, unsuccessful external intrusion, and successful external intrusion. Different levels of sabotage events correspond to different breeding risks and handling logics. Without real-time and accurate monitoring, it will lead to misjudgment of risk status and mismatch of handling actions, making it impossible to achieve precise control of the breeding scenario. Continuing the operation status of the sound and light-based expulsion actions in the previous steps, the location, movement, and existence status information of the intruding target are obtained in real time through point cloud monitoring equipment, infrared monitoring equipment, video monitoring equipment, and audio monitoring equipment deployed in the breeding area. At the same time, the status information of poultry, entry and exit status information, and environmental monitoring information in the breeding area are obtained. All kinds of information are compared with the three-level pre-set sabotage event judgment criteria in the previous steps one by one to accurately distinguish the existence status of the intruding target as continuous existence, temporary departure, or complete disappearance, and at the same time, accurately determine whether the three-level pre-set sabotage event has been triggered, as well as the specific level of sabotage event triggered.

[0220] For example, the system acquires various monitoring data in the breeding area in real time. If it detects poultry in a state of panic for only 3 seconds, accompanied by brief abnormal noise, and then returns to quiet within 5 seconds without repeated occurrences within 10 minutes, the system accurately determines that a Level 1 ordinary poultry panic event has been triggered. If it detects abnormal frequencies at the breeding entrance / exit, abnormal temperature detected by infrared, abnormal point clouds detected by video, accompanied by abnormal noise, but the intrusion target has not entered the core breeding area, the system accurately determines that a Level 2 external intrusion event has been triggered but the intrusion was unsuccessful. If it detects poultry exhibiting abnormally rapid gathering, dispersal, and movement, accompanied by abnormal noise, abnormal frequencies at the breeding entrance / exit, abnormal temperature detected by infrared, and abnormal point clouds detected by video, and the intrusion target has entered the core breeding area, the system accurately determines that a Level 3 external intrusion event has been triggered and the intrusion was successful. If none of the above abnormal states are detected, the system determines that no preset destructive event has been triggered.

[0221] This allows for precise monitoring of the presence of intruding targets during the implementation of sound and light-based deterrence actions, and the differentiation of three levels of pre-set destructive events. This avoids misjudgment of risk status and provides accurate and objective criteria for selecting subsequent graded response actions. It determines the selection logic and activation conditions of subsequent response actions. If a pre-set destructive event of the corresponding level is detected, the corresponding response action is initiated. If no pre-set destructive event is triggered, the monitoring status is maintained.

[0222] Different levels of preset disruptive events correspond to different levels of aquaculture risk. Response actions must be matched to the risk level and implemented in a tiered manner to avoid over- or under-response, which would prevent effective safety management of the aquaculture environment. Based on the monitoring results of previous steps, the system initiates corresponding response actions according to the three preset disruptive event levels. Level 1, ordinary panic in poultry, requires no additional response actions; only audible and visual deterrent actions and real-time monitoring are maintained, without triggering alarm information. Level 2, external intrusion but unsuccessful intrusion, triggers a warning-level response action, generating a warning-level alarm message and sending a warning notification to aquaculture management personnel, without activating physical defense facilities. Level 3, external intrusion and successful intrusion, triggers an alarm-level response action, generating an alarm-level alarm message and sending a detailed alarm notification to aquaculture management personnel, while simultaneously activating the linked defense actions of physical defense facilities.

[0223] For example, if the monitoring determines that a Level 1 ordinary poultry panic event has been triggered, the system will only continue to maintain strong light exposure and sound and light-based deterrent actions such as the zookeeper's voice announcement, continuously monitor the status in real time, without generating any alarm information or initiating any additional actions; if the monitoring determines that a Level 2 external intrusion event has been triggered but the intrusion was unsuccessful, the system will generate a warning-level alarm message containing the abnormal status and the location of the incident, and send it to the receiving terminal of the poultry management personnel, only providing a warning prompt, without activating physical defense facilities; if the monitoring determines that a Level 3 external intrusion event has been triggered and the intrusion was successful, the system will generate an alarm-level alarm message containing the intrusion status, poultry status, and risk level, and send it to the receiving terminal of the poultry management personnel, while simultaneously activating the physical defense facilities in the breeding area to implement coordinated defense actions of isolation and containment.

[0224] Based on the different levels of the three-level preset damage events, corresponding response actions are initiated to achieve graded response and precise control. This avoids excessive response actions that may affect the normal operation of aquaculture, or insufficient response actions that may fail to address the risks to aquaculture. It also provides a preliminary response status and operational logic basis for the termination of subsequent response actions after the disappearance of the intrusion target, and determines the timing and conditions for terminating subsequent response actions.

[0225] When the intrusion target disappears without triggering the corresponding level of preset damage event, or when the disappearance of the intrusion target only triggers the first level of ordinary poultry panic event, the handling action will be automatically terminated to avoid unnecessary handling actions from continuing, reduce system resource consumption, and restore the normal operation of the breeding area. The system monitors the existence status of the intrusion target in real time. When the monitoring determines that the intrusion target has completely left the breeding area, its existence status is completely disappeared, and only the first level of ordinary poultry panic event has been triggered, or no preset damage event has been triggered, the sound and light-based drive-away actions such as strong light illumination and the sound broadcast of the breeder will be automatically terminated immediately. All graded handling actions will be terminated simultaneously, no additional alarm information will be generated, no physical defense facility linkage will be activated, and the normal monitoring and operation status of the breeding area will be restored.

[0226] For example, if the intruding target completely leaves the breeding area after the sound and light-based deterrent action is implemented, only triggering a brief common poultry panic event and quickly recovering, the strong light exposure and the breeder's voice broadcast action will be automatically terminated immediately, and the normal monitoring status of the breeding area will be restored; if the intruding target disappears but triggers a second or third level preset damage event, the system will continue to maintain the corresponding level of handling action and will not terminate the handling action.

[0227] This allows all actions to be automatically terminated when the termination conditions are met, avoiding unnecessary consumption of system resources, quickly restoring the normal operation of the aquaculture area, maintaining the stability of aquaculture management, and automatically terminating the actions so that the aquaculture area can return to normal target monitoring status. This provides a stable operating basis for the continuous monitoring and judgment of moving targets in the aquaculture area, allowing for continued monitoring and judgment of moving targets.

[0228] Example 2:

[0229] like Figure 3 As shown in the figure, an intrusion alarm method for livestock sheds is provided in an embodiment of this application. The method includes:

[0230] The system acquires point cloud data and thermal imaging signals of moving targets inside the breeding house, as well as vibration signals at the inlet and outlet channels. After spatiotemporal registration, the data is fused to generate sensing data including point cloud features, thermal radiation gradient, and vibration spectrum.

[0231] Based on historically acquired sensory data, cluster analysis is used to extract and distinguish the activity areas, body size characteristics, and gait frequency characteristics of different types of poultry, so as to construct a spatiotemporal behavioral baseline that includes behavioral baseline characteristics and dynamic characteristics.

[0232] Extract moving targets from the current perception data and obtain multi-dimensional features of the moving targets. Perform hierarchical matching of the multi-dimensional features with the spatiotemporal behavior baseline. If the matching is successful, the moving target is determined to be a legitimate target. If the matching fails, the moving target is determined to be an intrusion target.

[0233] When a moving target is determined to be an intrusion target or an abnormal cloud of moving target points is detected, a tiered response is executed. If the intrusion target persists or a preset destructive event is triggered, an alarm message is sent and physical defense facilities are activated.

[0234] Since the principle of the method in this application embodiment is similar to that of the system described in this application embodiment, the implementation of the method is the same as that of the system, and the repeated parts will not be described again.

Claims

1. An intrusion alarm system for livestock sheds, characterized in that, include: The module consists of a perception module, a parsing module, a recognition module, and a response module. The sensing module is used to acquire point cloud data and thermal imaging signals of moving targets in the breeding house, as well as vibration signals at the inlet and outlet channels. After performing spatiotemporal registration, it is fused to generate sensing data including point cloud features, thermal radiation gradient and vibration spectrum. The parsing module is used to extract and distinguish the activity areas, body shape characteristics and gait frequency characteristics of different types of poultry based on historically acquired perception data through cluster analysis, and to construct a spatiotemporal behavioral baseline that includes behavioral baseline characteristics and dynamic characteristics. The identification module is used to extract moving targets from the current perception data and obtain the multi-dimensional features of the moving targets. The multi-dimensional features are then matched hierarchically with the spatiotemporal behavior baseline. If the match is successful, the moving target is determined to be a legitimate target; if the match fails, the moving target is determined to be an intrusion target. The response module is used to execute a graded response when the moving target is determined to be an intrusion target or when an abnormal cloud of moving target points is detected. When the intrusion target persists or a preset destructive event is triggered, alarm information is sent and physical defense facilities are activated.

2. The intrusion alarm system for livestock sheds as described in claim 1, characterized in that, The sensing data fused and generated in the sensing module is specifically used for: Acquire point cloud data, thermal imaging signals, and vibration signals at the inlet and outlet channels of moving targets within the breeding shed; and perform spatiotemporal registration of the point cloud data and thermal imaging signals based on the vibration signals. Using the moving target as a dynamic background, the system identifies the point cloud clusters of the moving target from the registered point cloud data and extracts the features of the point cloud clusters. It also extracts the thermal radiation gradient of the moving target from the registered thermal imaging signal and converts the vibration signal into the corresponding vibration spectrum. Using vibration spectrum as a link, the point cloud features of the same moving target are bound to thermal radiation gradient and fused to generate sensing data.

3. The intrusion alarm system for livestock sheds as described in claim 2, characterized in that, The spatiotemporal registration performed in the perception module is specifically used for: Time-domain peak detection is performed on vibration signals to screen vibration signals triggered by moving targets and determine vibration time-domain anchor points; Using the vibration time-domain anchor point as a reference, point cloud data frames and thermal imaging signal frames corresponding to the time window are extracted; Based on the propagation attenuation parameters of the vibration signal, the spatial coordinates of the moving target in the inlet and outlet channels of the breeding house are deduced. Using the spatial coordinates as the registration center, local spatial registration is performed on the point cloud data frame and the thermal imaging signal frame to complete the spatiotemporal registration.

4. The intrusion alarm system for livestock sheds as described in claim 3, characterized in that, The clustering analysis in the parsing module is specifically used for: The historically acquired sensory data is preprocessed, and invalid samples are removed based on the preset thermal radiation gradient threshold and gait frequency threshold in the poultry house, while candidate samples of poultry are retained. Using vibration spectrum and thermal radiation gradient as weighted features, density clustering was performed on candidate samples to separate different types of poultry samples. The body shape parameters, spatial coordinates, and gait frequency of the vibration spectrum of the corresponding point cloud features of poultry samples are extracted. After similarity clustering is performed, the activity area, body shape characteristics, and gait frequency characteristics of different types of poultry are obtained by combining the spatial zoning constraints of the breeding house. Based on the surface thermal radiation data of moving targets under different ambient temperatures, the thermal radiation gradient is fitted and dynamic characteristics of different types of poultry adapted to ambient temperature are constructed.

5. The intrusion alarm system for livestock sheds as described in claim 4, characterized in that, The similarity clustering performed in the parsing module is specifically used for: Calculate the similarity of body shape parameters and spatial coordinate distance of corresponding point cloud features in poultry samples, and cluster the poultry samples using Euclidean distance; Based on the spatial zoning constraints of the poultry house, the activity areas and body size characteristics of different types of poultry are matched; Gait frequencies corresponding to poultry samples are extracted, and the poultry samples are clustered using a density clustering algorithm. The consistency of gait frequencies in different clustering results is compared to obtain the gait frequency characteristics of different types of poultry.

6. The intrusion alarm system for livestock sheds as described in claim 5, characterized in that, The construction of the spatiotemporal behavior baseline in the parsing module is specifically used for: The activity area, body size characteristics, and gait frequency characteristics of different types of poultry were set as the behavioral benchmark characteristics of poultry. By binding the gait frequency characteristics of the same type of poultry with the corresponding activity area and body size characteristics, spatiotemporal association pairs are formed for each type of poultry. Based on historically acquired perception data, dynamic threshold ranges are set for the spatiotemporal correlation of each type of poultry, corresponding behavioral baseline features and dynamic features. For each type of poultry, a sub-baseline is established, which includes the corresponding behavioral baseline features, the dynamic features, the dynamic threshold range, and the spatiotemporal association rules. All sub-baselines are integrated to form a spatiotemporal behavioral baseline.

7. The intrusion alarm system for livestock sheds as described in claim 6, characterized in that, The multidimensional features of the moving target acquired in the recognition module are specifically used for: The current sensing data is preprocessed, and the core area of ​​the moving target is located based on the feature of the point cloud, with the moving target as the dynamic background. The thermal radiation gradient and vibration spectrum corresponding to the core area are then locked. Gait frequency features are extracted from the locked vibration spectrum, body shape parameters are extracted from the point cloud features, and gradient values ​​adapted to the current ambient temperature are extracted from the thermal radiation gradient as core features. Extract the state of the moving target point cloud and the real-time spatial coordinates of the moving target from the current sensing data as auxiliary features; The core features and auxiliary features are verified to determine whether they belong to the same moving target. Based on the gait frequency features, the verified core features and auxiliary features are correlated and normalized to form a multidimensional feature of the moving target.

8. The intrusion alarm system for livestock sheds as described in claim 7, characterized in that, The hierarchical matching is used for: The core features of the multidimensional features are compared with the behavioral benchmark features of the corresponding sub-baseline dimension by dimension to determine whether the core features are within the dynamic threshold range. If the core feature is not within the dynamic threshold range, the current sub-baseline is directly determined to be a failure to match, and the matching of other types of poultry sub-baselines is switched. If the core feature is within the dynamic threshold range, the auxiliary features of the multidimensional feature are compared with the behavioral baseline features, spatiotemporal association rules, and the dynamic features of the corresponding sub-baseline. If the auxiliary feature comparison is successful, then verify whether the multidimensional features involved in the matching are valid features of the same moving target; If the verification passes, the match is considered successful and the corresponding moving target is determined to be a legitimate target. If the verification fails or the auxiliary feature comparison is unqualified, the corresponding sub-baseline is considered to have failed to match. If all sub-baselines fail to match, the moving target is considered to be an intrusion target.

9. The intrusion alarm system for livestock sheds as described in claim 8, characterized in that, The response module is used for: When a moving target is determined to be an intrusion target or an anomaly is detected in the point cloud of a moving target, a graded response is immediately executed to carry out corresponding expulsion actions, and the existence status of the intrusion target and whether a preset destruction event is triggered are monitored in real time. If the intrusion target persists or a preset sabotage event is triggered, a corresponding alarm message will be generated and sent out, and the linkage defense action of the physical defense facilities will be activated at the same time. If the intrusion target disappears after being driven away and no preset destruction event is triggered, the driving action will automatically stop and no alarm information will be sent or physical defense facilities will be linked.

10. An intrusion alarm method for livestock sheds, implemented based on any one of claims 1-9, characterized in that, include: The system acquires point cloud data and thermal imaging signals of moving targets inside the breeding house, as well as vibration signals at the inlet and outlet channels. After spatiotemporal registration, the data is fused to generate sensing data including point cloud features, thermal radiation gradient, and vibration spectrum. Based on historically acquired sensory data, cluster analysis is used to extract and distinguish the activity areas, body size characteristics, and gait frequency characteristics of different types of poultry, so as to construct a spatiotemporal behavioral baseline that includes behavioral baseline characteristics and dynamic characteristics. Extract moving targets from the current perception data and obtain multi-dimensional features of the moving targets. Perform hierarchical matching of the multi-dimensional features with the spatiotemporal behavior baseline. If the matching is successful, the moving target is determined to be a legitimate target. If the matching fails, the moving target is determined to be an intrusion target. When a moving target is determined to be an intrusion target or an abnormal cloud of moving target points is detected, a tiered response is executed. If the intrusion target persists or a preset destructive event is triggered, an alarm message is sent and physical defense facilities are activated.

Images