Method and system for remotely identifying abnormal ducklings

Abstract

The invention relates to the technical field of rice field duck breeding, and discloses a method for remotely recognizing abnormal ducklings, and the method comprises the following steps: collecting the time sequence image data of a rice field duck group through an unmanned plane, and recognizing duck targets in each video frame based on a deep learning target detection model; extracting and quantifying the visibility mode characteristics of each duck individual in time sequence, wherein the visibility mode characteristics comprise the successfully detected frequency-visible frequency of the duck individual in unit time, and the continuous successfully tracked duration-continuous visible duration of each time; and micro-motion rhythm features of each duck individual in the visible time period are extracted and quantified. According to the method, automatic and high-precision identification of the abnormal ducklings in the rice field is realized through combination of unmanned aerial vehicle autonomous cruise and a deep learning technology, a traditional low-efficiency manual inspection mode is thoroughly innovated, and the monitoring efficiency is effectively improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of rice field duck raising, and in particular to a method and system for remotely identifying abnormal ducklings. BACKGROUND

[0002] Rice field duck raising is an important ecological agricultural model. Through the symbiotic relationship between duck groups and rice, ecological functions such as weeding, pest control, and fertilization are achieved, reducing the use of pesticides and fertilizers and improving the overall benefits of farmland. In actual breeding, 15-20 ducklings are usually raised per mu of rice field, and are placed in the field 1-2 weeks after transplanting.

[0003] However, ducklings placed in rice fields have low immunity and poor environmental adaptability, often exhibiting the following three types of abnormal conditions:

[0004] Sickness and death: ducklings are susceptible to infectious diseases such as duck cholera, duck infectious serositis, duck plague, and viral hepatitis. The carcasses of dead ducklings become a source of infection, leading to the spread of the epidemic;

[0005] Illness: diseases such as parasitic diseases and E. coli disease occur. If sick ducklings are not isolated and treated in time, they will become a source of infection for the entire duck group;

[0006] Injury: weak defense against natural enemies, vulnerable to attacks by natural enemies such as snakes, resulting in leg, foot, and wing injuries. If injured individuals are not treated in time, they will become infected.

[0007] Currently, the health of rice field duck groups is mainly monitored through manual patrols, which has the following significant drawbacks:

[0008] Low work efficiency: daily manual inspection of large areas of rice fields is time-consuming and labor-intensive;

[0009] Poor recognition accuracy: affected by the complex environment of rice fields (shading by seedlings, changes in light, water surface reflection, etc.), it is difficult for humans to accurately identify abnormal individuals;

[0010] Lack of real-time performance: continuous monitoring for 24 hours is not possible, and there is a lag in the detection of abnormalities;

[0011] Inadequate treatment: there is a long time interval between the detection of abnormalities and actual treatment, increasing the risk of cross-infection.

[0012] Therefore, there is an urgent need for a technical solution that can automatically, remotely, and accurately identify abnormal ducklings in rice fields, so as to timely detect and treat abnormal individuals, reduce duckling mortality, and improve the overall benefits of rice field duck raising. SUMMARY

[0013] To solve the technical problems presented in the background art, the present application provides a method and system for remotely identifying abnormal ducklings.

[0014] The application adopts the following technical scheme to realize the method for remotely identifying abnormal ducklings, which comprises the following steps:

[0015] Collecting time sequence image data of duck groups in rice fields by using a drone, and identifying duck targets in each video frame based on a deep learning target detection model;

[0016] Extracting and quantifying the visibility pattern features of each duck individual in the time sequence, wherein the visibility pattern features include the frequency of successful detection of the duck individual in unit time (frequency of visibility) and the duration of continuous successful tracking each time (continuous visibility duration);

[0017] Extracting and quantifying the micro-motion rhythm features of each duck individual in the visible period, wherein for each successfully detected duck target frame, the pixel-level light flow change amount is calculated, and periodicity or stationarity analysis is performed on the light flow change sequence in the continuous visible period, and the entropy value or dominant frequency of the light flow sequence is taken as the quantitative indicator of the micro-motion rhythm;

[0018] Constructing a normal behavior model of the duck group, wherein a classification model is trained using only the visibility pattern features and micro-motion rhythm features of the historical data of healthy ducks;

[0019] Inputting the features of the duck individual to be tested into the normal behavior model, and calculating the behavior pattern deviation degree thereof as a comprehensive abnormality score;

[0020] When the behavior pattern deviation degree of a duck individual exceeds an adaptive threshold value set based on the statistics of the healthy group, the duck is determined to be an abnormal individual.

[0021] Further, the collecting of time sequence image data of duck groups in rice fields by using a drone comprises:

[0022] Controlling the drone to fly at a flight height of 10-50 meters from the ground and a flight speed of 1-3 meters per second according to a preset grid path to cover the entire rice field area;

[0023] Obtaining high-definition image data at a collection rate of 1-5 frames per second, and the image resolution is not less than 1920x1080 pixels;

[0024] Performing preprocessing on the collected original image data, wherein the preprocessing includes denoising, color correction and image enhancement;

[0025] Arranging the preprocessed images in time sequence to generate a time sequence image sequence with time stamp.

[0026] Further, the identifying of duck targets in each video frame based on a deep learning target detection model comprises:

[0027] adopting a YOLOv5 model architecture as the deep learning object detection model;

[0028] training the model based on a special dataset containing duck bounding box annotations, wherein the training process uses a stochastic gradient descent optimizer and adopts a cosine annealing strategy to adjust the learning rate;

[0029] For the model output of each video frame, a non-maximum suppression algorithm is applied to eliminate overlapping detection boxes caused by occlusion or other factors, and the duck object detection result with the highest confidence is retained;

[0030] and

[0031] annotating the duck target bounding box for the rice field scene images collected by the unmanned aerial vehicle, and dividing the dataset into training set, validation set and test set in proportion, for model training, tuning and performance evaluation.

[0032] Further, the extraction and quantification of the visibility pattern features of each duck individual in time sequence include:

[0033] Based on the DeepSORT multi-target tracking algorithm, the duck targets detected in the continuous video frames are associated and tracked, and each duck individual is assigned and maintained a unique identity;

[0034] Within a preset sliding time window, the visibility frequency of each duck individual is calculated, which is the proportion of the number of frames successfully tracked to the total number of frames within the time window;

[0035] Extract all continuous visible periods of each duck individual within the time window and calculate the average continuous visible duration, which is the duration of continuous successful tracking.

[0036] Further, the visibility pattern features also include the maximum continuous visible duration and / or the standard deviation of the continuous visible duration, which are used to more comprehensively quantify the stability of duck individual visibility.

[0037] Further, the extraction and quantification of the micro-motion rhythm features of each duck individual in the visible period include:

[0038] For each successfully detected duck target box, the Farneback dense optical flow algorithm is used to calculate the pixel-level optical flow change in its region, and the average value of the optical flow amplitude of all pixel points in the region is taken as the global optical flow change representing the micro-motion intensity of the frame;

[0039] For a continuous visible period, the global optical flow change of each frame in the period is arranged in time sequence to construct an optical flow change sequence;

[0040] performing fast Fourier transform on the light flow change sequence, extracting a frequency component with the largest amplitude in a spectrum of the light flow change sequence as the dominant frequency, and using the dominant frequency to quantify periodic rhythm of the micro-movement.

[0041] Further, the extracting and quantifying the micro-movement rhythm feature further comprises:

[0042] performing information entropy calculation on the light flow change sequence to obtain an entropy value, and using the entropy value to quantify randomness or disorder degree of the micro-movement rhythm;

[0043] using the dominant frequency and the entropy value together as a quantification index of the micro-movement rhythm.

[0044] Further, the constructing the normal behavior model of the duck group comprises:

[0045] filtering and labeling behavior data containing only healthy duck individuals from historical data;

[0046] combining the visibility pattern feature and the micro-movement rhythm feature of each healthy duck individual into a comprehensive feature vector to construct a pure normal behavior data set;

[0047] using a one-class support vector machine algorithm to perform model training using the pure normal behavior data set as a training set, the one-class support vector machine defining a normal behavior boundary by finding a minimum hypersphere surrounding normal data in a feature space;

[0048] and

[0049] performing standardization preprocessing on features in the pure normal behavior data set before model training;

[0050] using a cross-validation method to optimize and select kernel function parameters and regularization parameters of the one-class support vector machine.

[0051] Further, the inputting the feature of the to-be-tested duck individual into the normal behavior model and calculating a behavior pattern deviation degree thereof as a comprehensive anomaly score comprises:

[0052] the behavior pattern deviation degree is quantified as a distance from a feature vector of the to-be-tested sample to a normal behavior boundary defined by the one-class support vector machine, and a greater distance value indicates a higher degree of deviation of the behavior pattern from the normal.

[0053] Further, the determining that a duck individual is an abnormal individual when a behavior pattern deviation degree of the duck individual exceeds an adaptive threshold value set based on healthy population statistics comprises:

[0054] calculating a baseline distribution of the deviation degree based on historical data of healthy ducks used to train the normal behavior model, and setting a predetermined high quantile value of the baseline distribution as an initial threshold value;

[0055] During the operation of the system, the threshold value is dynamically updated according to the deviation data of the duck individuals who are not recently judged as abnormal, so that the adaptive adjustment of the threshold value is realized;

[0056] The predetermined high quantile is a 95% quantile or a 99% quantile;

[0057] The period of dynamically updating the threshold value is daily or weekly.

[0058] The application further provides an abnormal duckling remote identification system suitable for a shielding environment, and the system comprises:

[0059] An image acquisition module is configured to acquire time sequence image data of a duck group in a rice field by using a UAV;

[0060] A target detection module is configured to identify duck targets in each video frame based on a deep learning target detection model;

[0061] A visibility pattern feature extraction module is configured to extract and quantify the visibility pattern features of each duck individual in time sequence, wherein the visibility pattern features comprise a frequency of successfully detected duck individuals in unit time, i.e., a visible frequency, and a duration of each continuous successful tracking, i.e., a continuous visible duration;

[0062] A micro-motion rhythm feature extraction module is configured to extract and quantify the micro-motion rhythm features of each duck individual in a visible period, wherein for each successfully detected duck target frame, the pixel-level light flow change amount is calculated, and periodicity or stationarity analysis is performed on the light flow change sequence in the continuous visible period, so that the entropy value or the dominant frequency of the light flow sequence is used as a quantitative index of the micro-motion rhythm;

[0063] A normal behavior model construction module is configured to train a classification model by using the visibility pattern features and the micro-motion rhythm features of only the historical data of healthy ducks;

[0064] An abnormal score calculation module is configured to input the features of a to-be-tested duck individual into the normal behavior model, and calculate the behavior pattern deviation degree of the to-be-tested duck individual as a comprehensive abnormal score;

[0065] An abnormality judgment module is configured to judge a duck individual as an abnormal individual when the behavior pattern deviation degree of the duck individual exceeds an adaptive threshold value set based on the statistics of a healthy group.

[0066] Further, the image acquisition module comprises:

[0067] A flight control unit is configured to control the UAV to fly at a flight height of 10-50 meters from the ground and a flight speed of 1-3 meters per second according to a preset grid path, so as to cover the entire rice field area;

[0068] An image acquisition unit is configured to acquire high-definition image data at a frame rate of 1-5 frames per second, with an image resolution of no less than 1920x1080 pixels.

[0069] A preprocessing unit is configured to preprocess the acquired raw image data, including denoising, color correction, and image enhancement.

[0070] A time sequence generation unit is configured to arrange the preprocessed images in chronological order to generate a time-stamped time sequence of images.

[0071] Further, the target detection module comprises:

[0072] A model architecture unit adopts a YOLOv5 model architecture as the deep learning target detection model.

[0073] A model training unit trains the model based on a special dataset containing duck bounding box annotations, wherein the training process uses a stochastic gradient descent optimizer and adopts a cosine annealing strategy to adjust the learning rate.

[0074] A detection box optimization unit is configured to apply a non-maximum suppression algorithm to the model output for each video frame to eliminate overlapping detection boxes caused by occlusion or other factors, and retain the duck target detection result with the highest confidence.

[0075] A dataset management unit is configured to annotate duck target bounding boxes for rice field scene images collected by a UAV, and divide the dataset into training, validation, and test sets in proportion for model training, tuning, and performance evaluation.

[0076] Further, the visibility pattern feature extraction module comprises:

[0077] A target tracking unit is configured to associate and maintain the identity of duck targets detected in consecutive video frames based on the DeepSORT multi-target tracking algorithm, and assign and maintain a unique identity for each duck individual.

[0078] A visible frequency calculation unit is configured to calculate the visible frequency of each duck individual within a preset sliding time window, which is the proportion of the number of successfully tracked frames to the total number of frames within the time window.

[0079] A continuous visible duration calculation unit is configured to extract all continuous visible periods of each duck individual within the time window and calculate the average continuous visible duration, which is the duration of a continuous successful tracking.

[0080] Further, the visibility pattern feature extraction module further comprises a stability quantification unit, configured to calculate the maximum continuous visibility duration and / or the standard deviation of the continuous visibility duration, for more comprehensively quantifying the stability of the individual duck visibility.

[0081] Further, the micro-motion rhythm feature extraction module comprises:

[0082] a light flow calculation unit, configured to, for each successfully detected duck target frame, calculate the light flow change of the pixel level in the region thereof by using a Farneback dense light flow algorithm, and take the average value of the light flow amplitudes of all the pixel points in the region as a global light flow change quantity representing the micro-motion intensity of the frame;

[0083] a sequence construction unit, configured to, for a continuous visibility period, arrange the global light flow change quantities of the frames in the period in time sequence to construct a light flow change sequence;

[0084] a periodicity analysis unit, configured to perform fast Fourier transform on the light flow change sequence, extract the frequency component with the largest amplitude in the frequency spectrum thereof as the dominant frequency, and use the dominant frequency to quantify the periodicity rhythm of the micro-motion.

[0085] Further, the micro-motion rhythm feature extraction module further comprises:

[0086] an entropy value calculation unit, configured to perform information entropy calculation on the light flow change sequence to obtain the entropy value, and use the entropy value to quantify the randomness or disorder degree of the micro-motion rhythm;

[0087] a feature combination unit, configured to jointly use the dominant frequency and the entropy value as the quantification indicators of the micro-motion rhythm.

[0088] Further, the normal behavior model construction module comprises:

[0089] a data screening unit, configured to screen and label the behavior data containing only healthy duck individuals from historical data;

[0090] a feature vector construction unit, configured to combine the visibility pattern features and the micro-motion rhythm features of each healthy duck individual into a comprehensive feature vector, and construct a pure normal behavior data set;

[0091] a model training unit, configured to use a one-class support vector machine algorithm to perform model training by taking the pure normal behavior data set as a training set, and the one-class support vector machine defines the normal behavior boundary by finding the smallest hyper-sphere surrounding the normal data in the feature space;

[0092] a preprocessing unit, configured to perform standardization preprocessing on the features in the pure normal behavior data set before model training;

[0093] The parameter optimization unit optimizes and selects kernel function parameters and regularization parameters of the one-class support vector machine by using a cross-validation method.

[0094] Further, the abnormality calculation module comprises a distance quantification unit configured to quantify the behavior pattern deviation as a distance from a feature vector of the to-be-tested sample to a normal behavior boundary defined by the one-class support vector machine, and a greater distance value indicates a higher degree of deviation of the behavior pattern from the normal.

[0095] Further, the abnormality determination module comprises:

[0096] A threshold setting unit configured to calculate a benchmark distribution of the deviation based on historical data of healthy ducks used to train the normal behavior model, and set a predetermined high quantile value of the benchmark distribution as an initial threshold value;

[0097] A threshold updating unit configured to dynamically update the threshold value based on deviation data of duck individuals that have not been determined as abnormal in recent days during system operation, so as to realize adaptive adjustment of the threshold value;

[0098] The predetermined high quantile value is a 95% quantile value or a 99% quantile value.

[0099] The period of dynamically updating the threshold value is daily or weekly.

[0100] Compared with the prior art, the present application has the following beneficial effects:

[0101] The present application realizes automatic and high-precision identification of abnormal ducklings in rice fields by combining unmanned aerial vehicle autonomous cruising with deep learning technology, completely innovates the traditional inefficient manual inspection mode, and effectively improves the monitoring efficiency. The present application innovatively integrates multi-dimensional behavior feature analysis such as visibility mode and micro-motion rhythm, and can still maintain a high abnormality recognition accuracy even in a complex environment blocked by seedlings. Early abnormality warning is realized by relying on a normal behavior model built by a one-class support vector machine. The entire system has a real-time response capability of minutes, and can push the warning in real time through the MQTT protocol, so that the management personnel can intervene in the first time. Further, the multi-modal feature analysis provided by the present application can accurately distinguish abnormal types such as dead ducks, sick ducks and disabled ducks, and is accurate in classification, thereby providing a basis for differentiated and accurate processing. The whole scheme of the present application has strong environmental adaptability, has a self-adaptive threshold mechanism, and greatly reduces the operation technical threshold by virtue of automatic color marking and intuitive color coding guidance of the unmanned aerial vehicle, so that ordinary breeders can also quickly locate and process abnormal individuals, thereby significantly reducing cross-infection and mortality of ducklings, and effectively improving the comprehensive benefits of rice duck farming. BRIEF DESCRIPTION OF DRAWINGS

[0102] Figure 1A method flow diagram for remotely identifying abnormal ducklings is provided for Embodiment 1 of the present application.

[0103] Figure 2 A flow diagram of the method for classifying abnormal duck individuals is provided for Embodiment 2 of the present application.

[0104] Figure 3 A flow diagram of the method for identifying and processing abnormal duck individuals in the field is provided for Embodiment 3 of the present application. DETAILED DESCRIPTION

[0105] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor fall within the scope of protection of the present application.

[0106] Embodiment 1: The present application provides a method for remotely identifying abnormal ducklings, the flow of which is shown in Figure 1 and includes the following steps:

[0107] S1, collecting time series image data of a duck group in a rice field by a drone, and identifying duck targets in each video frame based on a deep learning target detection model;

[0108] S2, extracting and quantifying the visibility pattern features of each duck individual in the time series;

[0109] S3, extracting and quantifying the micro-motion rhythm features of each duck individual in the visible period;

[0110] S4, constructing a normal behavior model of the duck group;

[0111] S5, inputting the features of the duck individual to be tested into the normal behavior model, and calculating the behavior pattern deviation degree as a comprehensive abnormality score;

[0112] S6, when the behavior pattern deviation degree of a duck individual exceeds the adaptive threshold value set based on the statistics of the healthy group, the duck is determined to be an abnormal individual.

[0113] In the present scheme, S1, time series image data of a duck group in a rice field is collected by a drone, and duck targets in each video frame are identified based on a deep learning target detection model, which is specifically implemented as:

[0114] When collecting time-series image data of duck groups in rice fields by drones, a DJI Matrice 350 RTK drone is used, equipped with a Zenmuse H20N hybrid sensor gimbal camera. The flight task is set as follows: flight height 30 meters above ground, flight speed 2 meters per second, flight according to the preset "bow" shaped grid path to ensure coverage of the entire rice field area. The camera acquires 1920x1080 pixel high-definition visible light video data at a rate of 3 frames per second.

[0115] The collected raw image data is preprocessed. First, denoising is performed using a Gaussian filter algorithm with a convolution kernel size of 5x5. Then color correction is performed based on the gray world assumption for white balance adjustment. The image details are enhanced using the contrast limited adaptive histogram equalization (CLAHE) algorithm, with a CLAHE block size of 8x8 and a contrast limit parameter of 2.0, highlighting the gray difference between the duck target and the background (such as the contrast between the duck white feathers and the rice field green seedlings). The preprocessed images are arranged in chronological order, with a millisecond-level timestamp (based on the UTC time synchronization of the drone) added to each frame of image, generating a time-series image sequence in the format "image file name_YYYYMMDD_HHMMSS_XXX.jpg" (XXX is the number of milliseconds), stored in a dedicated folder on the drone SD card, and simultaneously transmitted in real time to the ground data processing terminal for backup through the 4G / 5G module.

[0116] When identifying duck targets in each video frame based on a deep learning target detection model, the YOLOv5s model architecture is used. This model is trained using a dedicated dataset containing 5000 duck images collected in rice field scenarios, labeled with boundary boxes by experts using the LabelImg tool, and divided into training, validation and test sets in the ratio of 7:2:1. The model training uses a random gradient descent optimizer with an initial learning rate of 0.01, adjusted using a cosine annealing strategy, and a training period of 300 Epochs.

[0117] For each video frame, the non-maximum suppression algorithm is applied with an intersection over union threshold of 0.5 to eliminate overlapping detection boxes due to occlusion or other factors, and finally the highest confidence duck target detection result and its boundary box coordinates are retained.

[0118] S2, extract and quantify the visibility pattern features of each duck individual in time series, which is implemented as follows:

[0119] In the process of extracting the visibility pattern features, firstly, the identity association and trajectory maintenance of the detected duck targets in the continuous video frames are performed based on the DeepSORT multi-target tracking algorithm. The DeepSORT algorithm combines Kalman filtering to predict the target motion and extracts the CNN appearance features of the target for association matching, and assigns and maintains a unique identity ID (format such as "G001") for each duck individual.

[0120] In a preset sliding time window (window length T=60 seconds, sliding step S=10 seconds), the visibility frequency of each duck individual is calculated. The calculation formula of the visibility frequency is: "Visibility frequency = (the number of frames in which the target is successfully tracked in the window) / (the total number of frames in the window)".

[0121] At the same time, all continuous visible periods of each duck individual in the time window are extracted. A continuous visible period is defined as the time interval from the first successful tracking of the target to the next tracking loss. The continuous visible duration (unit: seconds) of each period is calculated, and the arithmetic mean of the durations of all periods in the window is taken to obtain the average continuous visible duration. To further quantify the visibility stability, the maximum continuous visible duration and the standard deviation of the continuous visible duration in the window are also calculated.

[0122] In this scheme, S3, the micro-motion rhythm features of each duck individual in the visible period are extracted and quantified, which is implemented as follows:

[0123] S3.1, definition of the region of interest and extraction of the frame sequence

[0124] For each duck target that is successfully detected and tracked, the region of interest of the frame is defined based on the bounding box coordinates of the duck in the current video frame.

[0125] From the start frame to the end frame of a continuous visible period (defined and provided by S2), the ROI corresponding to the duck identity ID in each frame is extracted in sequence to form a ROI image sequence. The sequence is strictly aligned with the continuous visible period.

[0126] S3.2, pixel-level optical flow field calculation

[0127] The Farneback dense optical flow algorithm is used to calculate the motion field between two consecutive frames (denoted as frame "t" and frame "t+1"). This algorithm models the neighborhood of each pixel point with a quadratic polynomial to calculate a dense and relatively smooth optical flow field.

[0128] The specific algorithm parameter settings are as follows:

[0129] Image pyramid layer number: "3". By constructing an image pyramid, a larger range of motion can be captured.

[0130] Scale per layer: “0.5”.

[0131] Average window size per layer: “15” pixels. This parameter determines the neighborhood size for the polynomial expansion.

[0132] Number of pyramid iterations per layer: “3”.

[0133] The algorithm outputs a two-channel optical flow matrix “Flow” of the same size as the original ROI image. Here, “Flow(x, y, 0)” represents the horizontal (x-axis) displacement component “dx” of pixel point “(x, y)”, and “Flow(x, y, 1)” represents the vertical (y-axis) displacement component “dy”. The unit is pixels per frame interval.

[0134] S3.3, Global Optical Flow Variation Calculation

[0135] To avoid interference from background or irrelevant pixels, a mask based on target detection confidence is first created, and only pixels in high-confidence areas are calculated.

[0136] For each valid pixel point within the ROI, calculate its optical flow vector magnitude (i.e., motion intensity):

[0137]

[0138] Then, calculate the arithmetic mean of the optical flow magnitudes of all valid pixels within the ROI, as the Global Optical Flow Variation (GOFV) of the duck individual in this frame time interval:

[0139]

[0140] is a scalar that quantifies the overall micro-motion intensity of the duck from frame “t” to frame “t+1”. An active and healthy duck may exhibit a moderate and regular “GOFV”, while a weak or abnormal duck may exhibit an excessively low (stationary) or high (agitated) and irregular “GOFV”.

[0141] S3.4, Optical Flow Variation Sequence Construction

[0142] For a continuous visible period of duration “N” frames, repeat S3.2 and S3.3 to obtain the “GOFV” value corresponding to each frame in this period (a total of “N-1” values, as optical flow calculation requires two frames).

[0143] Arrange these "GOFV" values ​​in chronological order to form a time series:

[0144]

[0145] This sequence constitutes the raw signal for analyzing micromotor rhythms.

[0146] S3.5, Time Series Analysis and Extraction of Rhythm Quantification Indicators

[0147] Next, signal analysis is performed on "GOFV_Sequence" to extract two core quantitative indicators: dominant frequency and entropy value, which respectively characterize the periodicity and regularity of the motion.

[0148] A. Dominant Frequency Calculation (Periodic Analysis)

[0149] 1. Preprocessing: Detrend the “GOFV_Sequence” by subtracting its linear fit value to eliminate slowly changing baseline drift.

[0150] 2. Fast Fourier Transform: Apply the FFT to the preprocessed sequence to transform it from the time domain to the frequency domain. The length of the FFT is set to be greater than or equal to the smallest power of 2 of the sequence length (for example, if the sequence length is 100, the FFT length is set to 128), and any insufficient length is padded with zeros.

[0151] 3. Frequency axis calculation: Frequency resolution "df = sampling frequency (Fs) / FFT length". Where the sampling frequency "Fs" is the video frame rate (e.g., 3 Hz).

[0152] 4. Spectrum Analysis: Calculate the magnitude of the FFT result to obtain the amplitude spectrum. Within the effective frequency range (typically 0 Hz to the Nyquist frequency "Fs / 2"), find the maximum peak value in the amplitude spectrum. The frequency corresponding to this peak value is the dominant frequency.

[0153] 5. Physical significance: The dominant frequency reflects the most prominent rhythmic component in the micro-movements of an individual duck. For example, healthy pecking and preening behaviors may exhibit a stable rhythm close to 1-2 Hz, while the lack of a dominant frequency may indicate disordered or aimless behavior.

[0154] B. Entropy calculation (stationarity / stochasticity analysis)

[0155] 1. Symbolization: To enhance robustness, the "GOFV_Sequence" is converted into a symbolic sequence. The mean "μ" and standard deviation "σ" of the sequence are calculated, and each "GOFV" value is converted into a symbol.

[0156] if The symbol is denoted as "0" (low motion).

[0157] if The symbol is denoted as "1" (medium motion).

[0158] if The symbol is denoted as "2" (high motion).

[0159] 2. Pattern (word) statistics: A sliding window (window length "L=3") is used to slide on the symbol sequence. Each time it moves one step, the occurrence count of all possible 3-symbol patterns (such as "000", "001", ..., "222") is recorded.

[0160] 3. Probability Distribution and Entropy Calculation: Divide the occurrence count of each pattern by the total number of patterns to obtain its empirical probability. Then, calculate the approximate entropy of the sequence (an entropy metric used for short sequences) using the information entropy formula, or directly calculate the Shannon entropy:

[0161] The summation range is all occurrences of pattern "i".

[0162] 4. Physical Significance: Entropy quantifies the uncertainty and randomness of micromotor patterns. Lower entropy indicates a more stable and regular movement rhythm (e.g., continuous pacing). Higher entropy indicates a more disordered and unpredictable movement rhythm (potentially a sign of stress or disease). Combined with the dominant frequency, low entropy with a clear dominant frequency usually represents healthy, rhythmic movement; high entropy with no dominant frequency represents abnormal, disordered movement.

[0163] S3.6, Feature Vector Generation

[0164] Ultimately, the micro-motion rhythm of a continuous visible time period is quantified into a two-dimensional feature vector: "[dominant frequency, entropy value]".

[0165] If there are multiple consecutive visible time periods within an analysis window, the feature vector is calculated for each time period, and then the average value of each dimension is taken as the final micro-movement rhythm feature of this duck individual within that window.

[0166] S4. Construct a normal behavior model for the duck flock, specifically implemented as follows:

[0167] When constructing a normal behavior model for a duck flock, livestock experts first selected all data from 50 confirmed healthy duck individuals based on video recordings from a month-long historical monitoring dataset.

[0168] For each analysis window of these healthy ducks (such as the 60-second sliding window defined in S2), the visibility pattern feature vector generated in S2 (including: visibility frequency, average continuous visibility duration, maximum continuous visibility duration, and standard deviation of continuous visibility duration) is concatenated with the micromotor rhythm feature vector generated in S3.6 (including: dominant frequency and entropy value) to form a six-dimensional comprehensive feature vector.

[0169] The combined feature vectors of all healthy ducks across all analysis windows are aggregated to construct a pure dataset of normal behavior.

[0170] A support vector machine (SVM) algorithm was employed, using all comprehensive feature vectors from the clean normal behavior dataset as the training set for model training. Before training, the dataset was Z-score standardized, calculating the mean (μ) and standard deviation (σ) for each feature dimension. A transformation of (x - μ) / σ was applied to all data, making the mean 0 and the standard deviation 1 for each feature dimension. The SVM used a radial basis function kernel, with its hyperparameter nu (representing the upper limit of the outlier proportion) and kernel bandwidth gamma optimized through 5-fold cross-validation, ultimately determining nu=0.05 and gamma=0.1. The model defines the normal behavior boundary by finding the smallest hypersphere in the feature space that encloses most of the normal data points.

[0171] After the model training is complete, save the model parameters and the mean (μ) and standard deviation (σ) required for standardization, for use in step S5.

[0172] In this scheme, S5, the characteristics of the duck to be tested are input into the normal behavior model, and its behavioral pattern deviation is calculated as a comprehensive abnormality score. Specifically, this is implemented as follows:

[0173] For the individual duck to be tested, its comprehensive feature vector is generated within the current analysis window following the same process as S2 and S3.6.

[0174] The target feature vector is then standardized using the same means (μ) and standard deviation (σ) saved in step S4. The standardized feature vector is then input into the pre-trained SVM model from step S4.

[0175] The deviation from the behavioral pattern is quantified as the signed distance from the standardized sample feature vector to the decision hyperplane (i.e., the normal behavior boundary) defined by a class of SVMs. A positive and larger distance value indicates that the sample deviates more from the normal behavior boundary, meaning the degree of behavioral pattern abnormality is higher. This deviation is a specific numerical value (e.g., +0.85).

[0176] In this scheme, S6, when the deviation of a duck's behavior pattern exceeds an adaptive threshold set based on statistics of a healthy population, the duck is determined to be an abnormal individual. Specifically, this is implemented as follows:

[0177] When setting the threshold, firstly, all samples from the normal behavior dataset used for training in S4 (i.e., the standardized comprehensive feature vectors of all healthy ducks) are input into a pre-trained SVM model to calculate the deviation of each healthy sample, thus forming a baseline distribution of the deviation of the healthy group. The 95th percentile of this baseline distribution is set as the initial threshold.

[0178] During system operation, to achieve adaptive adjustment of the threshold, the system automatically collects deviation data of all ducks that were not judged as abnormal in the past 24 hours every morning (i.e., the deviation values ​​calculated in step S5 for these samples considered "normal"). Based on this newly collected "normal" data, the distribution of deviation is recalculated, and the threshold is updated to the 95th percentile of the new distribution.

[0179] When the deviation of a duck individual calculated in real time (from the output of step S5) exceeds the currently effective adaptive threshold, the system immediately determines that the duck is an abnormal individual and generates an early warning signal. The early warning signal is sent to the breeding management APP via the MQTT protocol, and the information includes the abnormal individual ID, the abnormal time, the deviation value, and the extent to which it exceeds the threshold.

[0180] Example 2:

[0181] Building upon Example 1, this example further provides a method for classifying abnormal duck individuals, used to further categorize identified abnormal individuals into sick ducks, dead ducks, or disabled ducks. The process is as follows: Figure 2 As shown, it includes the following steps:

[0182] S7. Extract multimodal behavioral features from duck individuals identified as abnormal;

[0183] S8. Construct an anomaly type classification model;

[0184] S9. Input the multimodal behavioral features of the abnormal individuals to be classified into the abnormal type classification model, and output their respective abnormal types (sick duck / dead duck / disabled duck).

[0185] In this scheme, S7, for duck individuals determined to be abnormal, extract their multimodal behavioral features, specifically implemented as follows:

[0186] Based on the visibility pattern features and micro-motion rhythm features extracted in Example 1, the following four types of supplementary features are further extracted to form a 12-dimensional multimodal feature vector:

[0187] 1. Attitude stability characteristics (3D)

[0188] The lightweight pose estimation model MoveNet (a lightweight version of single pose) was used to detect 17 keypoints for each ROI image sequence, including keypoints for the head, neck, torso, and limbs.

[0189] The standard deviation of the positional changes of key points in the head over 10 consecutive frames (approximately 3.3 seconds) is calculated as an indicator of head stability.

[0190] Calculate the vertical positional variation of key trunk points (thoracic cavity and pelvis) as trunk undulation.

[0191] Calculate the average range of motion of the key points of the limbs relative to the trunk, as the limb range of motion.

[0192] 2. Motion trajectory characteristics (3D)

[0193] Based on DeepSORT tracking trajectory, the motion path within a 30-second time window is extracted.

[0194] Calculate the trajectory curvature: the ratio of the total path length to the straight-line distance from the starting point to the ending point.

[0195] Calculate the frequency of change in motion direction: the number of times the motion direction changes by more than 45° per unit time.

[0196] Calculate the concentration of the activity area: Use the convex hull algorithm to calculate the ratio of the trajectory coverage area to the area of ​​the minimum bounding rectangle.

[0197] 3. Appearance and morphological characteristics (3D)

[0198] Keyframes were extracted from the ROI sequence (one frame every 5 seconds), and deep features were extracted using a pre-trained ResNet18 model.

[0199] Calculate the pose symmetry score: based on the mirror symmetry evaluation of key points on the left and right sides.

[0200] Calculate profile compactness: the ratio of the area of ​​the target profile to the square of the profile perimeter.

[0201] Calculate feather neatness: Assess the degree of disorder in feather regions based on texture analysis.

[0202] 4. Behavioral Pattern Characteristics (2-dimensional)

[0203] Percentage of static periods: The percentage of periods within a 60-second analysis window where there is continuous stillness for more than 5 seconds.

[0204] Abnormal behavior frequency: The frequency of occurrence of detected uncommon behavioral patterns (such as spinning, head shaking, etc.).

[0205] Technical details of feature extraction:

[0206] Pose estimation: The MoveNet model in TensorFlow Lite format is used with an input size of 192×192, and inference is performed in real time on an embedded device.

[0207] Trajectory Analysis: The Douglas Peucker algorithm is used to simplify the motion trajectory while retaining the main motion features.

[0208] Appearance analysis: Using ResNet18 pre-trained on ImageNet, the last fully connected layer is removed, and a 512-dimensional feature vector is extracted.

[0209] Real-time performance guarantee: All feature extraction algorithms are optimized for embedded platforms, ensuring real-time operation on Jetson Nano-level devices.

[0210] In this solution, S8 involves constructing an anomaly type classification model, specifically implemented as follows:

[0211] Collect a dataset of outlier samples from historical data that have been accurately labeled by livestock experts:

[0212] 200 diseased duck samples, covering symptoms of common diseases such as avian influenza and duck plague.

[0213] The sample of dead ducks consisted of 150 ducks, representing various states from immediate death to some time after death.

[0214] 150 disabled duck samples, including those with leg and foot disabilities, wing injuries, etc.

[0215] Model selection and training:

[0216] Gradient boosting decision tree (LightGBM) is used as the classifier because it performs well on tabular data and has a fast inference speed.

[0217] The five-fold cross-validation method was used for hyperparameter optimization. The main parameters adjusted were: number of leaf nodes: 31; learning rate: 0.05; feature sampling ratio: 0.8; L1 regularization coefficient: 0.1.

[0218] Before training, feature importance analysis is performed, and features with importance below 1% are removed.

[0219] The SMOTE algorithm is used to oversample the classes with fewer samples to ensure class balance.

[0220] Model performance metrics: Achieve the following performance on the test set:

[0221] Overall accuracy: 92.3%.

[0222] Accuracy rate for identifying dead ducks: 96.7%.

[0223] Accuracy rate for identifying sick ducks: 89.5%.

[0224] The accuracy rate for identifying disabled ducks was 90.8%.

[0225] In this scheme, S9, the multimodal behavioral features of the individual to be classified as an anomaly are input into the anomaly type classification model, and the anomaly type to which it belongs is output. Specifically, this is implemented as follows:

[0226] For individuals identified as abnormal in real time, the following classification process is executed:

[0227] 1. Feature extraction: Extract 12-dimensional multimodal feature vectors according to step S7.

[0228] 2. Feature preprocessing: Normalize features using standardized parameters saved during training.

[0229] 3. Model Inference: Input the processed features into the trained LightGBM classification model.

[0230] 4. Confidence assessment: Calculate the predicted probability for each category and set a confidence threshold of 0.7.

[0231] If the highest probability is ≥0.7, output the corresponding category directly.

[0232] If the highest probability is less than 0.7, it is marked as "pending confirmation" and requires manual review.

[0233] 5. Time-series consistency check: For the same abnormal individual, the classification results of three consecutive analysis windows are subject to majority voting to ensure classification stability.

[0234] Classification decision rules:

[0235] Characteristics for identifying dead ducks:

[0236] Micromotor rhythm: The dominant frequency is close to 0Hz, and the entropy value is extremely low.

[0237] Postural stability: Head stability index <0.1 pixels, limb range of motion <0.5 pixels.

[0238] Motion trajectory: The trajectory curvature is close to 1, and the frequency of change in motion direction is 0.

[0239] Characteristics for identifying sick ducks:

[0240] Micromotor rhythm: The dominant frequency is abnormal (deviating from the healthy range of 12Hz), and the entropy value is high.

[0241] Behavioral patterns: The proportion of inactive periods is >60%, and the frequency of abnormal behaviors increases.

[0242] Appearance: Feather neatness score is low, and posture symmetry is reduced.

[0243] Characteristics for identifying disabled ducks:

[0244] Motion trajectory: The trajectory curvature is greater than 2.5, and the change in motion direction follows a fixed pattern.

[0245] Postural stability: Abnormal trunk undulation and asymmetrical limb range of motion.

[0246] Behavioral pattern: The activity area is highly concentrated, exhibiting limited movement.

[0247] Example 3:

[0248] Building upon Examples 1 and 2, this example further provides a method for on-site identification and handling of abnormal ducks based on drone-based color-marking, facilitating rapid identification and handling of abnormal individuals by poultry farmers. The process is as follows: Figure 3 As shown, it includes the following steps:

[0249] In this scheme, S10, a differentiated color-marking scheme is formulated based on the anomaly classification results;

[0250] S11. Control the drone to automatically spray color markings on identified abnormal individuals;

[0251] S12. Provide farmers with simple handling guidelines based on color coding.

[0252] In this plan, S10, a differentiated color-marking scheme is formulated based on the anomaly classification results, and the specific implementation is as follows:

[0253] Based on the anomaly type classification results output in step S9 of Example 2, the system automatically selects the corresponding color spraying marking scheme:

[0254] Dead Duck Marking Procedure: Marking Color: Red; Spray Dosage: 3 ml (ensure full coverage); Marking Location: Spray covering the entire body, centered on the torso; Urgency Level: High (requires immediate treatment).

[0255] Marking procedure for sick ducks: Marking color: yellow; Spray dosage: 1.5 ml (to form obvious spots); Marking location: central area of ​​the back; Treatment urgency: medium (must be treated on the same day).

[0256] Disability marking protocol for ducks: Marking color: blue; Spray dosage: 1.5 ml (to form a noticeable spot); Marking location: central area of ​​the back; Treatment urgency: low (to be treated as appropriate).

[0257] The spray paint uses environmentally friendly, biodegradable, food-grade pigments, comprising: 70% water-soluble carrier, 25% food coloring, and 5% adhesive. This formula is safe and harmless to ducks and can completely degrade within 7 days in a natural environment.

[0258] In this scheme, S11, the control drone automatically sprays color markings on identified abnormal individuals, specifically as follows:

[0259] S11.1 Structure and Control of the Color Spraying Device

[0260] The drone is equipped with a dedicated three-color spraying system, which includes: three independent pigment reservoirs (red, yellow, and blue), each with a capacity of 500 ml; a miniature peristaltic pump with a flow control accuracy of ±0.1 ml; a solenoid valve control switch with a response time of <10 milliseconds; and a fan-shaped nozzle with a spray angle of 60 degrees, which can form a marking area with a diameter of approximately 0.8 meters at a height of 5 meters. The spraying system communicates with the drone's flight control system via a CAN bus, receiving spraying commands and providing status information.

[0261] S11.2 Automatic Color Spraying Positioning and Control

[0262] Once the system identifies an abnormal individual and determines the color spraying plan, the following automatic color spraying process is executed:

[0263] Precise target positioning: The drone hovers 5 meters above the target duck and locks the target position based on a template matching algorithm using real-time video stream, with a positioning error of less than 0.2 meters.

[0264] Motion trajectory prediction: The Kalman filter algorithm is used to predict the target's motion trajectory within the next 2 seconds and calculate the lead time for spraying color.

[0265] Spraying parameter calculation: Adjust nozzle pressure according to target distance: set pressure to 0.1MPa at a distance of 5 meters; Adjust spraying timing according to target movement speed: trigger 0.5-1.0 seconds in advance for moving targets; Control spraying dosage according to marking scheme: achieved by controlling the peristaltic pump working time (1 ml / s).

[0266] Color spraying execution and verification:

[0267] Trigger the solenoid valve at the optimal time to execute the color spraying;

[0268] After the paint is applied, the marking image is captured by a camera.

[0269] Use a color recognition algorithm to verify the marking effect;

[0270] If marking fails, the re-spraying process will be executed automatically (up to 2 re-sprays).

[0271] S11.3 Multi-target marker sequence optimization

[0272] When multiple abnormal individuals are identified simultaneously, the system uses a greedy algorithm to optimize the marking path: first, mark dead ducks (red marker); then mark sick ducks (yellow marker); and finally mark disabled ducks (blue marker). The path planning ensures that the drone travels the shortest distance and the marking efficiency is the highest.

[0273] In this plan, S12 provides farmers with simple processing guidelines based on color coding, specifically implemented as follows:

[0274] S12.1 Color Recognition Training Materials

[0275] Provide simple training materials for aquaculture personnel:

[0276] Color identification cards: displaying red, yellow, and blue colors and their corresponding meanings.

[0277] Example of actual product photos: showing the actual marking effect.

[0278] Processing priority description: Red > Yellow > Blue

[0279] S12.2 On-site positioning guidance scheme

[0280] Drone hovering indication: After completing the color marking, the drone hovers at a height of 10-15 meters above the target area and flashes the corresponding color via onboard LED lights (red light indicates dead duck, yellow light indicates sick duck, and blue light indicates disabled duck).

[0281] Voice alert broadcast: The drone will repeatedly broadcast voice prompts, such as "A dead duck marked with a red mark has been found. Please handle it immediately."

[0282] Batch processing mode: When multiple abnormal individuals exist in the same area, the drone hovers over the area to instruct the farmers to handle them in one go.

[0283] S12.3, Configuration of Dedicated Processing Tools

[0284] Each type of abnormal individual is equipped with a special handling tool: red marking (dead ducks): equipped with long-handled pick-up clips and special sealing bags; yellow marking (sick ducks): equipped with isolation nets and transfer cages; blue marking (disabled ducks): equipped with ordinary pick-up clips and care cages; all tools are managed by color coding to facilitate quick identification and use by farmers.

[0285] S12.4 Processing Progress Tracking

[0286] After handling each abnormal individual, the farmer waves to the drone. The drone uses its onboard vision system to recognize the wave, records the processing progress, and automatically returns to base after all abnormal individuals have been dealt with.

[0287] This embodiment greatly reduces the technical barrier to entry by using color-coded markings and intuitive on-site guidance, enabling farmers with limited education to effectively identify and handle abnormal ducks, thereby improving the efficiency of breeding management.

[0288] All calculations involved in the embodiments are dimensionless numerical calculations, and the preset parameters and thresholds in the calculations are set by those skilled in the art according to the actual situation.

[0289] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, in the form of a computer program product.

[0290] Those skilled in the art will recognize that the modules and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and inventive constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0291] In addition, the functional modules in the various embodiments of this application can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module.

[0292] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of modules is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple modules or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or modules may be electrical, mechanical, or other forms.

[0293] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0294] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for remotely identifying abnormal ducklings, characterized in that, The method includes the following steps: By collecting time-series image data of duck flocks in rice paddies using drones, and identifying duck targets in each video frame based on a deep learning target detection model; Extract and quantify the temporal visibility pattern features of each duck individual: the visibility pattern features include the frequency of successful detection of the duck individual per unit time - visibility frequency, and the duration of each consecutive successful tracking - continuous visibility duration; Extract and quantify the micromotor rhythm features of each duck individual during the visible period: For each successfully detected duck target box, calculate its pixel-level optical flow change, and perform periodic or stationary analysis on the optical flow change sequence during the continuous visible period, using the entropy value or dominant frequency of the optical flow sequence as the quantitative index of micromotor rhythm. Constructing a normal behavior model for a duck flock: A classification model is trained using the visibility pattern features and micromotor rhythm features that contain only historical data of healthy ducks; The characteristics of the ducks to be tested are input into the normal behavior model, and the deviation of their behavior patterns is calculated as a comprehensive abnormality score. When the deviation of a duck's behavior pattern exceeds an adaptive threshold set based on statistics of a healthy population, the duck is determined to be an abnormal individual.

2. The method according to claim 1, characterized in that, The time-series image data of duck flocks in rice paddies collected by drones includes: Control the drone to fly at an altitude of 10-50 meters above the ground and a speed of 1-3 meters per second, following a preset grid pattern to cover the entire rice paddy area; Acquire high-definition image data at a capture rate of 1-5 frames per second, with an image resolution of no less than 1920x1080 pixels; The acquired raw image data is preprocessed, including denoising, color correction, and image enhancement. The preprocessed images are arranged in chronological order to generate a time-series image sequence with timestamps.

3. The method according to claim 1, characterized in that, The method for identifying duck targets in each video frame using a deep learning-based object detection model includes: The YOLOv5 model architecture was adopted as the deep learning object detection model. The model is trained on a dedicated dataset containing duck bounding box annotations, wherein the training process uses a stochastic gradient descent optimizer and a cosine annealing strategy to adjust the learning rate. For the model output of each video frame, a non-maximum suppression algorithm is applied to eliminate overlapping detection boxes caused by occlusion or other factors, and retain the duck target detection result with the highest confidence. and Duck target bounding boxes were labeled on rice paddy scene images collected by drones, and the dataset was divided into training set, validation set and test set according to the proportion for model training, optimization and performance evaluation.

4. The method according to claim 1, characterized in that, The extraction and quantification of the temporal visibility pattern features of each individual duck includes: Based on the DeepSORT multi-target tracking algorithm, the duck targets detected in continuous video frames are associated with identities and their trajectories are maintained, and a unique identity is assigned and maintained for each individual duck. Within a preset sliding time window, the visibility frequency of each duck individual is calculated, where the visibility frequency is the proportion of frames in which the individual is successfully tracked within the time window to the total number of frames. Extract all consecutive visible periods for each duck individual within the time window and calculate its average consecutive visible duration, which is the duration of a single successful tracking.

5. The method according to claim 4, characterized in that, The visibility pattern features also include the maximum continuous visibility duration and / or the standard deviation of continuous visibility duration, used to more comprehensively quantify the stability of individual duck visibility.

6. The method according to claim 1, characterized in that, The extraction and quantification of the micromotor rhythm characteristics of each individual duck during the visible time period includes: For each successfully detected duck bounding box, the Farneback dense optical flow algorithm is used to calculate the pixel-level optical flow changes within its region, and the average value of the optical flow amplitude of all pixels within the region is used as the global optical flow change representing the micro-motion intensity of that frame. For a continuous visible time period, the global optical flow changes of each frame within that time period are arranged in chronological order to construct an optical flow change sequence; The optical flow variation sequence is subjected to a fast Fourier transform, and the frequency component with the largest amplitude in its spectrum is extracted as the dominant frequency, which is used to quantify the periodic rhythm of micro-motion.

7. The method according to claim 6, characterized in that, The extraction and quantification of micro-motor rhythm features also includes: Information entropy is calculated on the optical flow change sequence to obtain the entropy value, which is used to quantify the randomness or disorder of the micro-motion rhythm; The dominant frequency and the entropy value are used together as quantitative indicators of the micro-motor rhythm.

8. The method according to claim 1, characterized in that, The construction of the normal behavior model of the duck flock includes: Filter and label behavioral data containing only healthy duck individuals from historical data; The visibility pattern features and micromotor rhythm features of each healthy duck individual are combined into a comprehensive feature vector to construct a pure normal behavior dataset; A support vector machine algorithm is used to train the model using the pure normal behavior dataset as the training set. The support vector machine defines the normal behavior boundary by finding the smallest hypersphere that encloses the normal data in the feature space. and Before model training, the features in the pure normal behavior dataset are standardized and preprocessed. Cross-validation is used to optimize the selection of kernel function parameters and regularization parameters for a class of support vector machines.

9. The method according to claim 1 or 8, characterized in that, The step of inputting the characteristics of the duck individual to be tested into the normal behavior model and calculating its behavioral pattern deviation as a comprehensive anomaly score includes: The deviation of the behavior pattern is quantified as the distance from the feature vector of the sample to be tested to the normal behavior boundary defined by the support vector machine. The larger the distance value, the higher the degree of deviation of the behavior pattern from the normal.

10. The method according to claim 9, characterized in that, The determination that a duck is an abnormal individual when its behavioral pattern deviates beyond an adaptive threshold set based on statistics of a healthy population includes: Based on the historical data of healthy ducks used to train the normal behavior model, the baseline distribution of its deviation is calculated, and the predetermined high quantile value of the baseline distribution is set as the initial threshold. During system operation, the threshold is dynamically updated periodically based on the deviation data of ducks that have not been judged as abnormal recently, so as to achieve adaptive adjustment of the threshold. The predetermined higher quantile is the 95th percentile or the 99th percentile; The dynamic update threshold is updated daily or weekly.

Images