A method for determining abnormal retention behavior of fish population in a sturgeon culture pond

Abstract

The application discloses a kind of fishery abnormal retention behavior determination methods for sturgeon breeding pond, comprising the following steps: step S1: obtain the sonar image of fish school by multiple imaging sonar equipment, generate panoramic sonar graph;Step S2: the adaptive threshold segmentation is carried out to the panoramic sonar graph based on echo intensity statistical characteristics, and fish school target area is extracted;Step S3: based on the spatial structure characteristics of sonar echo, extract pool bottom echo boundary line and divide bottom layer area;Step S4: the bottom layer distribution density and movement intensity of fish school are calculated;Step S5: the bottom layer retention time of fish school is calculated;Step S6: bottom layer distribution density, movement intensity and retention time are fused to construct abnormal determination model, realize fish school abnormal retention state detection.The application is suitable for the behavior monitoring of industrialized sturgeon culture, can provide accurate data support for sturgeon health management, effectively improve the detection rate of sturgeon abnormal behavior, provide reliable technical support for intelligent aquaculture system.

Description

Technical Field

[0001] This invention relates to the field of intelligent monitoring technology in aquaculture, and in particular to a method for determining abnormal lingering behavior of fish in sturgeon farming ponds, applicable to behavioral monitoring in industrialized sturgeon farming scenarios. Background Technology

[0002] Sturgeon are typical benthic fish, usually inhabiting the bottom areas of aquaculture water. However, when sturgeon experience abnormal conditions such as hypoxia, disease, or death, they often remain at the bottom of the pond for extended periods and exhibit reduced activity. Traditional monitoring methods based on visible light images suffer significant image quality degradation during rainy seasons or in turbid water conditions, making effective monitoring difficult. While existing sonar equipment can detect fish schools, it is mostly used for determining the presence of fish or overall counting, lacking sophisticated modeling of fish school behavior and methods for identifying abnormal states, especially lacking dedicated monitoring mechanisms for the benthic characteristics of sturgeon.

[0003] In the prior art, CN120182326A discloses a method for tracking fish schools and calculating swimming speed based on sonar images. However, this tracking method struggles to achieve accurate individual segmentation and stable ID tracking when dealing with dense schools of fish, failing to meet the monitoring needs of high-density bottom-feeding behavior in sturgeon farming ponds. CN114637014B discloses a fish school behavior recognition system and method based on an underwater robot in unmanned fish farms, which calculates an activity coefficient after feeding to determine the abnormality rate. However, this method relies on intermittent feeding stimuli and is affected by the degree of hunger in the fish school, making it impossible to achieve continuous, all-weather monitoring of abnormal bottom-feeding behavior. CN118711138B discloses a deep-sea aquaculture monitoring method and system based on sonar images, utilizing sonar technology to identify and statistically analyze changes in fish school growth and distribution. However, it is mainly applicable to open marine environments and is not suitable for aquaculture ponds. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of the prior art and provide a method for determining abnormal lingering behavior of sturgeon in sturgeon farming ponds. Through multi-imaging sonar collaborative observation, acoustic-shadow topology stripping, and multi-dimensional feature fusion, it can achieve stable identification of abnormal lingering behavior of sturgeon at the bottom of the pond under turbid water conditions.

[0005] The objective of this invention is achieved through the following technical solution: a method for determining abnormal lingering behavior in sturgeon farming ponds, comprising the following steps:

[0006] Step S1: Acquire sonar images of the fish school using multiple imaging sonar devices installed on the side wall of the aquaculture pond, and generate a panoramic sonar map;

[0007] Step S2: Based on the statistical characteristics of echo intensity, perform adaptive threshold segmentation on the panoramic sonar image to extract the target area of ​​the fish school;

[0008] Step S3: Based on the spatial structure characteristics of sonar echoes, extract the bottom echo boundary line and divide the bottom layer region through the correlation analysis of echo intensity gradient and acoustic shadow characteristics;

[0009] Step S4: Calculate the bottom distribution density and movement intensity of the fish population;

[0010] Step S5: Adaptively adjust parameter thresholds based on historical sonar density time series data, and calculate the fish swarm's bottom dwell time based on the sliding time window and interruption tolerance time;

[0011] Step S6: Integrate the underlying distribution density, movement intensity, and dwell time to construct an anomaly detection model and realize the detection of abnormal dwelling state of fish groups.

[0012] Furthermore, in step S1, multiple imaging sonar devices installed on the sidewall of the aquaculture pond acquire sonar images of the fish school to generate a panoramic sonar map. The specific steps are as follows:

[0013] Step S101: Install at equal intervals on the sidewalls of the circular aquaculture pond High-frequency imaging sonar ( Each sonar unit is fixed underwater by a support frame. Depth, and the sonar main axis is perpendicular to the horizontal plane. The pitch angle points towards the center of the pool bottom;

[0014] Step S102: Synchronous Acquisition The path contains raw sonar stream data with overlapping fields of view, i.e., sonar images of fish schools;

[0015] Step S103: Establish a unified horizontal projection coordinate system based on the geometric model of the aquaculture pond to generate a panoramic sonar image covering the entire pond.

[0016] Further, step S2 performs adaptive threshold segmentation on the panoramic sonar image based on echo intensity statistical features to extract the target area of ​​the fish school, specifically including:

[0017] Step S201: Perform statistical analysis on the echo intensity of the panoramic sonar image, and calculate the average echo intensity and intensity dispersion of the current frame;

[0018] The average echo intensity is the arithmetic mean of the grayscale values ​​of all pixels in the panoramic sonar image, and the intensity dispersion is the global standard deviation.

[0019] Step S202: Perform adaptive threshold segmentation based on average echo intensity and intensity dispersion to obtain the initial target region;

[0020] Specifically, the average echo intensity is used as a benchmark, and a dynamic offset is set according to the intensity dispersion to generate a threshold matrix of the same size as the image. This threshold matrix is ​​then used to perform pixel-by-pixel binarization of the panoramic sonar image to obtain the initial target region. When the grayscale value of a pixel is higher than the adaptive threshold at the corresponding location, it is determined to be a fish target; otherwise, it is determined to be background.

[0021] The dynamic offset is set by calculating the global standard deviation of the panoramic sonar image echo intensity. As a measure of intensity dispersion, a correction coefficient c is set, and the dynamic offset is... ;

[0022] The threshold matrix is ​​generated by using the average echo intensity. Based on the baseline, a threshold matrix is ​​generated by combining dynamic offsets. ,in To create a threshold matrix with the same spatial dimension as the panoramic sonar image, the value of each element in the matrix is ​​equal to... To support pixel-by-pixel parallel binarization operations.

[0023] Step S203: Perform connected component analysis on the initial target region and remove isolated noise regions with an area smaller than a preset threshold;

[0024] Step S204: Perform morphological processing on the preserved area to obtain a continuous and stable target area for the fish school.

[0025] Furthermore, due to the bottom-dwelling biological characteristics of sturgeon, this invention identifies and separates bottom-dwelling fish groups by recognizing the topological relationship between high-intensity echo blocks (fish bodies) and the acoustic shadowing area behind them, thereby reducing the interference of fish aggregation on bottom echo observation and avoiding errors in bottom depth determination. This invention does not precisely reconstruct the actual pond bottom, but rather uses fish body interference removal, temporal filtering, and historical information compensation methods to stably estimate the bottom echo boundary, thus constructing a reliable reference benchmark for bottom region division. Step S3, based on the spatial structural characteristics of sonar echoes, extracts the bottom echo boundary line and divides the bottom region through correlation analysis of echo intensity gradient and acoustic shadow characteristics, including steps S301 to S303:

[0026] Step S301: Based on the correlation analysis between echo intensity gradient and acoustic shadow features, identify and remove fish targets close to the bottom of the pool to remove interference, and extract continuous bottom echo boundary lines.

[0027] (1) Echo feature point extraction: Extract the one-dimensional echo signal of the j-th sound beam from the panoramic sonar image, and calculate the echo intensity of the j-th sound beam along the radial direction. gradient magnitude Set gradient threshold , will satisfy Furthermore, sampling points whose echo intensity falls within a preset range are marked as candidate feature points;

[0028] The gradient magnitude The calculation formula is:

[0029]

[0030] in, For the j-th sound beam at a radial distance The echo intensity value at that location, The echo intensity value of the previous sampling point; the panoramic sonar map is composed of multiple beam scan lines at different azimuth angles. The j-th beam refers to the one-dimensional radial echo signal extracted along the j-th azimuth angle in the panoramic sonar map. The candidate feature points are marked by traversing all azimuth angles in sequence.

[0031] (2) Sound shadow feature detection: Starting from the candidate feature points, search outward along the direction of sound beam propagation for continuous low-energy regions with intensity lower than the environmental reference threshold, define them as associated sound shadow regions, and record their radial length. ;

[0032] (3) Fish bottom peeling determination: Define the strong echo region from the start of the candidate feature point to the associated acoustic shadow region, and calculate the radial length of the region. Related sound and shadow length The ratio, if If the signal falls within the preset range of sturgeon biological characteristics, it is determined to be a fish target; otherwise, it is determined to be a valid echo from the bottom of the pond.

[0033] Step S302: Use time-domain median filtering to calibrate the boundary line of the pool bottom echo to eliminate boundary jumps caused by fish blocking the view;

[0034] Specifically, the bottom echo boundary data of the pool in the current frame and several consecutive previous frames are acquired to form a time window. The echo distance values ​​of the same beam angle within the time window are sorted, and the median value is selected as the estimated value of the bottom boundary at the current moment to suppress instantaneous abnormal fluctuations caused by fish interference.

[0035] Step S303: Using the calibrated bottom echo boundary as the reference plane, construct a spatial layer of preset thickness upwards to determine the bottom layer region.

[0036] Furthermore, unlike traditional time-based filtering methods, this invention drives state transitions through the underlying retention frequency and combines interruption tolerance and retention decay mechanisms to achieve segmented dynamic modeling of fish school behavior. Step S4 calculates the underlying distribution density and movement intensity of the fish school, and its sub-steps include:

[0037] Step S401: Count the number of target pixels of the fish in the bottom area, and calculate the bottom distribution density of the fish in the bottom area in combination with the total number of pixels in the bottom area;

[0038] The formula for calculating the density of fish at the bottom of a school is:

[0039]

[0040] in, for The number of pixels occupied by the fish target identified in the underlying region at any given time; The total number of pixels in the bottom layer region is given; by calculating the ratio of these two values, the distribution density of the fish in the bottom layer space is obtained. ;

[0041] Step S402: Calculate the fish movement intensity based on the echo differential of adjacent frame panoramic sonar images;

[0042] The motion intensity is used to characterize the total pixel change in the target area of ​​the fish school per unit time, and its calculation formula is as follows:

[0043]

[0044] in, For the time of change; This is the bottom layer region; for The binary representation of the fish target in the bottom region at any given time is set to 1 if it belongs to the fish target, and 0 otherwise. It represents the change in the presence state of the fish swarm at that pixel location.

[0045] Step S403: Perform time series smoothing on the density and motion intensity to reduce interference caused by instantaneous fluctuations.

[0046] Furthermore, step S5, which adaptively adjusts the parameter threshold based on historical sonar density time-series data and calculates the fish swarm's bottom dwell time based on a sliding time window and an interruption tolerance time, includes the following sub-steps:

[0047] Step S501: Based on historical sonar density time series data, identify the behavior pattern category at the current moment through unsupervised clustering analysis, and adjust the parameters adaptively accordingly;

[0048] (1) Historical density time series data acquisition and preprocessing;

[0049] Collect time-series data of fish swarm density at the bottom layer, output from step S4, over a preset number of days, with the sampling interval synchronized with the sonar frame rate. Resample the daily density time-series data according to a sliding time window, and extract the density mean, density standard deviation, and time-coded features within each window to construct a behavioral feature vector.

[0050] (2) Unsupervised cluster analysis;

[0051] The behavioral feature vectors are clustered using a Gaussian mixture model. The number of clusters K is automatically determined by the Bayesian information criterion, resulting in multiple behavioral patterns, including high bottom-probability windows, high bottom-probability windows, and transition windows. A corresponding set of tolerance parameters is preset for each behavioral pattern, and a mapping relationship between categories and parameters is established.

[0052] (3) Adaptive parameter mapping and online update.

[0053] Based on the posterior probability of the feature vector belonging to each window at the current time, select the window category corresponding to the highest probability, and map to obtain the tolerance parameter set {staying frequency threshold} corresponding to that category. Interruption tolerance time threshold Retention time retention coefficient Residence time decay coefficient A weighted average method is used to update cluster centers and model parameters to adapt to the evolution of fish circadian rhythms.

[0054] Step S502: Based on the sliding time window, the frequency of fish staying in the bottom area is counted, and the fish are determined to be in a staying state or an interrupted state by comparing the staying frequency with the threshold.

[0055] Set length as A sliding time window is used to calculate the percentage of frames that satisfy the underlying density condition within that window. The calculation formula is as follows:

[0056]

[0057] in, To satisfy within the window Frame count, This represents the distribution density of the fish population in the bottom-layer space in the current frame. The minimum threshold that is set; This represents the total number of frames within the time window.

[0058] According to the dwell frequency With the dwell frequency threshold By comparing the fish population, we can determine whether the fish are lingering at the bottom of the pond: when When, it is determined to be in a state of stagnation; when When this occurs, it is determined to be in an interrupt state.

[0059] Step S503: Perform the corresponding time accumulation and tolerance processing according to the stagnation state or interruption state.

[0060] Unlike traditional methods based on instantaneous threshold judgment, this invention proposes a modeling mechanism for dwell behavior based on state transition and time memory. By using an asymmetric update strategy for interruption tolerance time and dwell time, it achieves robust suppression of short-term disturbances.

[0061] Regarding the dwell time in the state of being in a ... Accumulate and clear the accumulated interruption time: ; ;

[0062] In the interrupted state, the interruption time Accumulate and compare the interruption time with the interruption tolerance time threshold. Relationship: ;when At the same time, the length of stay is retained and accumulated. ;when At that time, the dwell time is attenuated. ;in The retention time retention factor is used. The two coefficients are obtained through step S501 to form an asymmetric time response mechanism, which is the residence time decay coefficient.

[0063] The mechanism is used to suppress misjudgments of state caused by short-term fish detachment from the bottom, thereby improving the stability of abnormal lingering detection.

[0064] Furthermore, step S6 integrates the underlying distribution density, movement intensity, and dwell time to construct an anomaly detection model, thereby detecting abnormal dwelling states of fish schools. This includes the following sub-steps:

[0065] Step S601: Obtain the bottom distribution density, movement intensity and dwell time of the fish school, and obtain the score of abnormal dwelling state of the fish school based on the weighted calculation model;

[0066] The scoring formula for abnormal stay status is as follows:

[0067]

[0068] in, The density of fish at the bottom of the school. Exercise intensity The normalized value, For normalized Duration of stay;

[0069] Step S602: Compare the abnormal stay status score with a preset threshold;

[0070] Step S603: When the abnormal stagnation status score exceeds the threshold, it is determined that the fish group is in an abnormal stagnation status.

[0071] The beneficial effects of this invention are: by using multi-channel sonar collaborative acquisition, data fusion and projection transformation, and accurate identification of the bottom area, this invention integrates multi-dimensional features such as bottom distribution density, movement intensity and residence time to construct an anomaly judgment model, thereby achieving stable identification of abnormal residence behavior of sturgeon under turbid water conditions, effectively improving the detection rate of abnormal behavior and reducing the false alarm rate caused by short-term bottom-leaving behavior. Attached Figure Description

[0072] Figure 1 This is a flowchart of the present invention.

[0073] Figure 2 This is a diagram showing the sonar deployment.

[0074] Figure 3 This is a schematic diagram of the geometric relationship of sonar elevation projection.

[0075] Figure 4 This is a schematic diagram of the coordinate transformation of the sonar projection plane.

[0076] Figure 5 This is a diagram for determining the bottom layer's retention status.

[0077] Figure 6 Logic diagram for calculating the time fish spend at the bottom of the fish swarm. Detailed Implementation

[0078] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings, but the scope of protection of the present invention is not limited to the following description.

[0079] like Figure 1 As shown, the purpose of this invention is to provide a method for determining abnormal lingering behavior in sturgeon farming ponds. Through multi-sonar collaborative observation and spatiotemporal feature modeling of fish schools, it achieves stable identification of abnormal lingering behavior at the bottom of fish schools under turbid water conditions. Figure 1 This includes the following steps:

[0080] Step S1: Acquire sonar images of the fish school using multiple imaging sonar devices installed on the side wall of the aquaculture pond, and generate a panoramic sonar map;

[0081] Step S2: Based on the statistical characteristics of echo intensity, perform adaptive threshold segmentation on the panoramic sonar image to extract the target area of ​​the fish school;

[0082] Step S3: Based on the spatial structure characteristics of sonar echoes, extract the bottom echo boundary line and divide the bottom layer region through the correlation analysis of echo intensity gradient and acoustic shadow characteristics;

[0083] Step S4: Calculate the bottom distribution density and movement intensity of the fish population;

[0084] Step S5: Adaptively adjust parameter thresholds based on historical sonar density time series data, and calculate the fish swarm's bottom dwell time based on the sliding time window and interruption tolerance time;

[0085] Step S6: Integrate the underlying distribution density, movement intensity, and dwell time to construct an anomaly detection model and realize the detection of abnormal dwelling state of fish groups.

[0086] Furthermore, in step S1, multiple imaging sonar devices installed on the sidewall of the aquaculture pond acquire sonar images of the fish school to generate a panoramic sonar map. The specific steps are as follows:

[0087] Step S101: Install at equal intervals on the sidewalls of the circular aquaculture pond High-frequency imaging sonar ( Each sonar unit is fixed underwater by a support frame. Depth, and the sonar main axis is perpendicular to the horizontal plane. The pitch angle points towards the center of the pool bottom;

[0088] Figure 2 The diagram shows the sonar layout. In the embodiments of this application, three sonars are installed.

[0089] Step S102: Synchronous Acquisition The path contains raw sonar stream data with overlapping fields of view;

[0090] Step S103: Establish a unified horizontal projection coordinate system based on the geometric model of the aquaculture pond to generate a panoramic sonar image covering the entire pond.

[0091] The core of step S103, panoramic sonar image generation, lies in transforming the raw sonar data in the polar coordinate system to a horizontal plane in the Cartesian coordinate system using a spatial projection algorithm, and then performing multi-path fusion.

[0092] (1) Based on the installation depth of each sonar device and the pitch angle relative to the horizontal plane Spatial geometric correction is performed on the slant range data in the original sonar image to convert the sonar slant range spatial data into horizontal projection spatial data.

[0093] (2) Establish a unified Cartesian coordinate system with the geometric center of the aquaculture pond as the origin. Based on the installation position and orientation angle of each sonar device on the pond wall, perform coordinate transformation on the geometrically corrected image data and map it to the unified coordinate system.

[0094] like Figure 3 Sonar equipment at elevation angle Emit sound waves and measure the slant distance of the target point. And project this point onto the horizontal plane, the horizontal projection distance is... , The calculation formula is as follows:

[0095]

[0096] in, This represents the slant range corresponding to the sonar echo. Sonar equipment pitch angle;

[0097] like Figure 4 The horizontal projection distance and azimuth information of sonar images in polar coordinates are converted into local rectangular coordinates, and the conversion relationship is as follows:

[0098]

[0099]

[0100] in, The azimuth angle of the sonar beam relative to the vertical axis of the local coordinate system; , These are the horizontal and vertical coordinates in a local rectangular coordinate system established with the sonar equipment as the origin;

[0101] Based on this, combined with the first The installation position and orientation angle of each sonar device in the aquaculture pond are used to perform rotation and translation transformations on the local coordinates to obtain the global coordinates. The transformation relationship is as follows:

[0102]

[0103] in, For the first The orientation angle of a sonar device, that is, the angle between the sonar main axis direction and the vertical axis of the global coordinate system; , The first The horizontal and vertical coordinates of each sonar device in the global coordinate system; , These are the x and y coordinates in the transformed global coordinate system, respectively;

[0104] (3) Identify the overlapping areas of adjacent road sonar fields of view, perform weighted fusion processing on the overlapping areas, and generate a panoramic sonar map;

[0105] The weighted fusion formula is as follows, in which the fusion weights gradually change linearly along the splicing direction:

[0106]

[0107] in, To integrate the panoramic sonar image in coordinates Pixel intensity at that location; and The two adjacent sonar images participating in the fusion are in the same coordinate system. Pixel intensity at that location; and These are the corresponding fusion weight coefficients. .

[0108] Further, step S2 performs adaptive threshold segmentation on the panoramic sonar image based on echo intensity statistical features to extract the target area of ​​the fish school, specifically including:

[0109] Step S201: Perform statistical analysis on the echo intensity of the panoramic sonar image, and calculate the average echo intensity and intensity dispersion of the current frame;

[0110] Step S202: Perform adaptive threshold segmentation based on average echo intensity and intensity dispersion to obtain the initial target region;

[0111] Specifically, the average echo intensity is used as a benchmark, and a dynamic offset is set according to the intensity dispersion to generate a threshold matrix of the same size as the image. This threshold matrix is ​​then used to perform pixel-by-pixel binarization of the panoramic sonar image to obtain the initial target region. When the grayscale value of a pixel is higher than the adaptive threshold at the corresponding location, it is determined to be a fish target; otherwise, it is determined to be background.

[0112] The dynamic offset is set by calculating the global standard deviation of the panoramic sonar image echo intensity. As a measure of intensity dispersion, a correction coefficient c is set (0.3 in this example), and the dynamic offset is... ;

[0113] The threshold matrix is ​​generated by using the average echo intensity. Based on the baseline, a threshold matrix is ​​generated by combining dynamic offsets. ,in To create a threshold matrix with the same spatial dimension as the panoramic sonar image, the value of each element in the matrix is ​​equal to... To support pixel-by-pixel parallel binarization operations.

[0114] Step S203: Perform connected component analysis on the initial target region and remove isolated noise regions with an area smaller than a preset threshold;

[0115] Step S204: Perform morphological processing on the preserved area to obtain a continuous and stable target area for the fish school.

[0116] The morphological processing includes closing operations to fill the holes inside the target, and temporal filtering to smooth the target boundary.

[0117] Further, step S3, based on the spatial structural characteristics of the sonar echo, extracts the bottom echo boundary line and divides the bottom layer region through correlation analysis of echo intensity gradient and acoustic shadow characteristics, including steps S301 to S303:

[0118] Step S301: Based on the correlation analysis between echo intensity gradient and acoustic shadow features, identify and remove fish targets close to the bottom of the pool to remove interference, and extract continuous bottom echo boundary lines.

[0119] (1) Echo feature point extraction: Extract the one-dimensional echo signal of the j-th sound beam from the panoramic sonar image, and calculate the echo intensity of the j-th sound beam along the radial direction. gradient magnitude Set gradient threshold , will satisfy Furthermore, sampling points whose echo intensity falls within a preset range are marked as candidate feature points;

[0120] The gradient magnitude The calculation formula is:

[0121]

[0122] in, For the j-th sound beam at a radial distance The echo intensity value at that location, The echo intensity value of the previous sampling point; the panoramic sonar map is composed of multiple beam scan lines at different azimuth angles. The j-th beam refers to the one-dimensional radial echo signal extracted along the j-th azimuth angle in the panoramic sonar map. The candidate feature points are marked by traversing all azimuth angles in sequence.

[0123] (2) Sound shadow feature detection: Starting from the candidate feature points, search outward along the direction of sound beam propagation for continuous low-energy regions with an intensity lower than the environmental reference threshold (45 in the embodiment of this application), define them as associated sound shadow regions, and record their radial length. ;

[0124] (3) Fish bottom peeling determination: Define the strong echo region from the start of the candidate feature point to the associated acoustic shadow region, and calculate the radial length of the region. Related sound and shadow length The ratio, if If the signal falls within the preset range of sturgeon biological characteristics (1.2 in the embodiments of this application), it is determined to be a fish target; otherwise, it is determined to be a valid echo from the bottom of the pond.

[0125] like Figure 5 As shown, when the sonar beam sweeps across a sturgeon lying on the bottom, the surface of the fish (high-intensity echo patch) will first reflect the sound waves, and then a sound shadow area will appear behind the fish. Figure 5In the bottom-feeding sturgeon echo curve at point (a), the downward-concave curve represents the fish target interference. The core logic of this step identifies this concave shape feature (high-intensity echo + rear acoustic shadow) and then removes it.

[0126] Step S302: Use time-domain median filtering to calibrate the boundary line of the pool bottom echo to eliminate boundary jumps caused by fish blocking the view;

[0127] Specifically, the bottom echo boundary data of the current frame and several consecutive previous frames are acquired to form a time window (the most recent 5 frames are selected in the embodiment of this application). The echo distance values ​​of the same beam angle within the time window are sorted, and the median value is selected as the estimated value of the bottom boundary at the current moment to suppress instantaneous abnormal fluctuations caused by fish interference.

[0128] Step S303: Using the calibrated bottom echo boundary as the reference plane, construct a spatial layer of preset thickness upwards to determine the bottom layer region.

[0129] In this embodiment, the preset thickness is dynamically adjusted according to the water depth of the aquaculture pond, and in this embodiment of the application, it is taken as 30% of the pond depth.

[0130] Further, step S4 calculates the bottom distribution density and movement intensity of the fish population, including the following sub-steps:

[0131] Step S401: Count the number of target pixels of the fish in the bottom area, and calculate the bottom distribution density of the fish in the bottom area in combination with the total number of pixels in the bottom area;

[0132] The formula for calculating the density of fish at the bottom of a school is:

[0133]

[0134] in, for The number of pixels occupied by the fish target identified in the underlying region at any given time; This represents the total number of pixels in the bottom layer region.

[0135] Step S402: Calculate the fish movement intensity based on the echo differential of adjacent frame panoramic sonar images;

[0136] The motion intensity is used to characterize the total pixel change in the target area of ​​the fish school per unit time, and its calculation formula is as follows:

[0137]

[0138] in, The change time is 1 second in the embodiments of this application. This is the bottom layer region; for The binary representation of the fish target in the bottom region at any given time is set to 1 if it belongs to the fish target, and 0 otherwise. It represents the change in the presence state of the fish swarm at that pixel location.

[0139] Step S403: Perform time series smoothing on the density and motion intensity to reduce interference caused by instantaneous fluctuations.

[0140] Specifically, the distribution density is calculated using a sliding window averaging algorithm over multiple consecutive frames. and exercise intensity Filtering is performed. By introducing weights from historical frame data (5 frames in this embodiment), transient fluctuations are reduced, resulting in a smooth feature sequence that reflects the trend of fish behavior.

[0141] Furthermore, step S5, which adaptively adjusts the parameter threshold based on historical sonar density time-series data and calculates the fish swarm's bottom dwell time based on a sliding time window and an interruption tolerance time, includes the following sub-steps:

[0142] Step S501: Based on historical sonar density time series data, identify the behavior pattern category at the current moment through unsupervised clustering analysis, and adjust the parameters adaptively accordingly;

[0143] (1): Historical density time series data acquisition and preprocessing;

[0144] Collect time-series data of fish swarm density at the bottom layer, output from step S4, over a preset number of days, with the sampling interval synchronized with the sonar frame rate. Resample the daily density time-series data according to a sliding time window, and extract the density mean, density standard deviation, and time-coded features within each window to construct a behavioral feature vector.

[0145] (2): Unsupervised clustering analysis;

[0146] The behavioral feature vectors are clustered using a Gaussian mixture model. The number of clusters K is automatically determined by the Bayesian information criterion, resulting in multiple behavioral patterns, including high bottom-probability windows, high bottom-probability windows, and transition windows. A corresponding set of tolerance parameters is preset for each behavioral pattern, and a mapping relationship between categories and parameters is established.

[0147] (3): Adaptive parameter mapping and online update.

[0148] Based on the posterior probability of the feature vector belonging to each window at the current time, select the window category corresponding to the highest probability, and map to obtain the tolerance parameter set {staying frequency threshold} corresponding to that category. Interruption tolerance time threshold Retention time retention coefficient Residence time decay coefficient A weighted average method is used to update cluster centers and model parameters to adapt to the evolution of fish circadian rhythms.

[0149] In the embodiments of this application, the weighting formula is:

[0150]

[0151] in, This represents the updated cluster centers or parameter model; This represents the cluster centers or parameter model calculated at the current moment; Represents the cluster centers or parametric model of the previous time step;

[0152] Step S502: Based on the sliding time window, the frequency of fish staying in the bottom area is counted, and the fish are determined to be in a staying state or an interrupted state by comparing the staying frequency with the threshold.

[0153] Set length as A sliding time window (60 seconds in the embodiments of this application) is used to calculate the percentage of frames that meet the underlying density conditions within the window. The calculation formula is as follows:

[0154]

[0155] in, To satisfy within the window Frame count, This represents the distribution density of the fish population in the bottom-layer space in the current frame. The minimum threshold is set (40% in the embodiments of this application; if more than 40% of the bottom area is occupied, it means that most of the fish are crowded here, which is an abnormal clustering). This represents the total number of frames within the time window.

[0156] According to the dwell frequency With the dwell frequency threshold By comparing the fish population, we can determine whether the fish are lingering at the bottom of the pond: when When, it is determined to be in a state of stagnation; when When this occurs, it is determined to be in an interrupt state.

[0157] Step S503: Perform the corresponding time accumulation and tolerance processing according to the stagnation state or interruption state.

[0158] Its core logic is as follows Figure 6 During the state of being in a suspended state, the suspension time is... Accumulate and clear the accumulated interruption time: ; ;

[0159] In the interrupted state, the interruption time Accumulate and compare the interruption time with the interruption tolerance time threshold. Relationship: ;when At the same time, the length of stay is retained and accumulated. ;when At that time, the dwell time is attenuated. ;in The retention time retention factor is used. This is the residence time decay coefficient.

[0160] Furthermore, step S6 integrates the underlying distribution density, movement intensity, and dwell time to construct an anomaly detection model, thereby detecting abnormal dwelling states of fish schools. This includes the following sub-steps:

[0161] Step S601: Obtain the bottom distribution density, movement intensity and dwell time of the fish school, and obtain the score of abnormal dwelling state of the fish school based on the weighted calculation model;

[0162] The scoring formula for abnormal stay status is as follows:

[0163]

[0164] in, The density of fish at the bottom of the school. Exercise intensity The normalized value, For normalized Residence time. In the embodiments of this application, the maximum value normalization method is used for processing.

[0165] Since the dwell time has a significant impact on the behavior of sturgeon when considering abnormal behavior, this embodiment of the application takes... =0.2, =0.2, =0.6.

[0166] Step S602: Compare the abnormal stay status score with a preset threshold;

[0167] Step S603: When the abnormal lingering state is scored When the threshold is exceeded (0.7 in the embodiments of this application), the fish are determined to be in an abnormal stagnation state.

[0168] The foregoing description illustrates and describes a preferred embodiment of the present invention. However, as previously stated, it should be understood that the present invention is not limited to the forms disclosed herein and should not be construed as excluding other embodiments. It can be used in various other combinations, modifications, and environments, and can be altered within the scope of the inventive concept described herein through the foregoing teachings or techniques or knowledge in related fields. Any modifications and variations made by those skilled in the art that do not depart from the spirit and scope of the present invention should be within the protection scope of the appended claims.

Claims

1. A method for determining abnormal lingering behavior in sturgeon farming ponds, characterized in that: Includes the following steps: Step S1: Acquire sonar images of the fish school using multiple imaging sonar devices installed on the side wall of the aquaculture pond, and generate a panoramic sonar map; Step S2: Based on the statistical characteristics of echo intensity, perform adaptive threshold segmentation on the panoramic sonar image to extract the target area of ​​the fish school; Step S3: Based on the spatial structure characteristics of sonar echoes, extract the bottom echo boundary line and divide the bottom layer region through the correlation analysis of echo intensity gradient and acoustic shadow characteristics; Step S4: Calculate the bottom distribution density and movement intensity of the fish population; Step S4 includes the following sub-steps: Step S401: Count the number of target pixels of the fish in the bottom area, and calculate the bottom distribution density of the fish in the bottom area in combination with the total number of pixels in the bottom area; The formula for calculating the density of fish at the bottom of a school is: ； in, for The number of pixels occupied by the fish target identified in the underlying region at any given time; This represents the total number of pixels in the bottom layer region. Step S402: Calculate the fish movement intensity based on the echo differential of adjacent frame panoramic sonar images; The motion intensity is used to characterize the total pixel change in the target area of ​​the fish school per unit time, and its calculation formula is as follows: ； in, For the time of change; This is the bottom layer region; for The binary representation of the fish target in the bottom region at any given time is set to 1 if it belongs to the fish target, and 0 otherwise. This represents the change in the state of the fish group at that pixel location; Step S403: Perform time-series smoothing on the bottom distribution density and movement intensity of the fish population; Step S5: Adaptively adjust parameter thresholds based on historical sonar density time series data, and calculate the fish swarm's bottom dwell time based on the sliding time window and interruption tolerance time; Step S6: Integrate the underlying distribution density, movement intensity, and dwell time to construct an anomaly detection model and realize the detection of abnormal dwelling state of fish groups.

2. The method for determining abnormal lingering behavior of fish in sturgeon farming ponds according to claim 1, characterized in that: Step S1 includes the following sub-steps: Step S101: Install at equal intervals on the sidewalls of the circular aquaculture pond High-frequency imaging sonar Each sonar unit is fixed underwater by a support frame. Depth, and the sonar main axis is perpendicular to the horizontal plane. The pitch angle points towards the center of the pool bottom; Step S102: Synchronous Acquisition The path contains raw sonar stream data with overlapping fields of view, i.e., sonar images of fish schools; Step S103: Establish a unified horizontal projection coordinate system based on the geometric model of the aquaculture pond to generate a panoramic sonar image covering the entire pond.

3. The method for determining abnormal lingering behavior of fish in sturgeon farming ponds according to claim 1, characterized in that: Step S2 includes the following sub-steps: Step S201: Perform statistical analysis on the echo intensity of the panoramic sonar image, and calculate the average echo intensity and intensity dispersion of the current frame; Step S202: Perform adaptive threshold segmentation based on average echo intensity and intensity dispersion to obtain the initial target region; Step S203: Perform connected component analysis on the initial target region and remove isolated noise regions with an area smaller than a preset threshold; Step S204: Perform morphological processing on the preserved area to obtain a continuous and stable target area for the fish school.

4. The method for determining abnormal lingering behavior of fish in sturgeon farming ponds according to claim 2, characterized in that: Step S3 includes the following sub-steps: Step S301: Based on the correlation analysis between echo intensity gradient and acoustic shadow features, identify and remove fish targets close to the bottom of the pool to remove interference, and extract continuous bottom echo boundary lines. Step S302: Use time-domain median filtering to calibrate the boundary line of the pool bottom echo to eliminate boundary jumps caused by fish blocking the view; Step S303: Using the calibrated bottom echo boundary as the reference plane, construct a spatial layer of preset thickness upwards to determine the bottom layer region.

5. The method for determining abnormal lingering behavior of fish in sturgeon farming ponds according to claim 4, characterized in that: Step S301 includes: Echo feature point extraction: Extract the one-dimensional echo signal of the j-th sound beam from the panoramic sonar image, and calculate the echo intensity of the j-th sound beam along the radial direction. gradient magnitude Set gradient threshold , will satisfy Furthermore, sampling points whose echo intensity falls within a preset range are marked as candidate feature points; The gradient magnitude The calculation formula is: ； in, For the j-th sound beam at a radial distance Echo intensity value at that location The echo intensity value of the previous sampling point; the panoramic sonar map is composed of multiple beam scan lines at different azimuth angles. The j-th beam refers to the one-dimensional radial echo signal extracted along the j-th azimuth angle in the panoramic sonar map. The candidate feature points are marked by traversing all azimuth angles in sequence. Sound shadow feature detection starts from the candidate feature points and searches outward along the sound beam propagation direction for continuous low-energy regions with intensity below the environmental reference threshold. These regions are defined as associated sound shadow regions, and their radial lengths are recorded. ; Fish bottom stripping determination defines the strong echo region from the starting point of the candidate feature point to the associated acoustic shadow region, and calculates the radial length of this region. Related sound and shadow length The ratio, if If the signal falls within the preset range of sturgeon biological characteristics, it is determined to be a fish target; otherwise, it is determined to be a valid echo from the bottom of the pond.

6. The method for determining abnormal lingering behavior of fish in sturgeon farming ponds according to claim 1, characterized in that: Step S5 includes the following sub-steps: Step S501: Based on historical sonar density time series data, identify the behavior pattern category at the current moment through unsupervised clustering analysis, and adjust the parameters adaptively accordingly; Step S502: Based on the sliding time window, the frequency of fish staying in the bottom area is counted, and the fish are determined to be in a staying state or an interrupted state by comparing the staying frequency with the threshold. Set length as A sliding time window is used to calculate the percentage of frames that satisfy the underlying density condition within that window. The calculation formula is as follows: ； in, To satisfy within the window Frame count, This represents the distribution density of the fish population in the bottom-layer space in the current frame. The minimum threshold that is set; This represents the total number of frames within the time window. According to the dwell frequency With the dwell frequency threshold By comparing the fish population, we can determine whether the fish are lingering at the bottom of the pond: when When, it is determined to be in a state of stagnation; when When this occurs, it is determined to be in an interrupt state; Step S503: Based on the described stagnation or interruption state, perform the corresponding time accumulation and tolerance processing; While in a state of confinement, the duration of confinement is... Accumulate and clear the accumulated interruption time: ; ; When in an interrupted state, the interruption time is... Accumulate and compare the interruption time with the interruption tolerance time threshold. Relationship; ;when At the same time, the length of stay is retained and accumulated. ;when At that time, the dwell time is attenuated. ;in The retention time retention factor is used. The two coefficients are obtained through step S501 to form an asymmetric time response mechanism, which is the residence time decay coefficient.

7. The method for determining abnormal lingering behavior of fish in sturgeon farming ponds according to claim 6, characterized in that: Step S501 includes: Historical density time series data acquisition and preprocessing; Collect time-series data of fish swarm bottom distribution density output by step S4 within a preset number of days, with the sampling interval synchronized with the sonar frame rate; resample the daily density time-series data according to a sliding time window, extract the density mean, density standard deviation and time coding features in each window, and construct a behavioral feature vector; Unsupervised clustering analysis; The behavioral feature vectors are clustered using a Gaussian mixture model. The number of clusters K is automatically determined by the Bayesian information criterion to obtain multiple behavioral patterns. A corresponding set of tolerance parameters is preset for each behavioral pattern to establish a mapping relationship between categories and parameters. Adaptive parameter mapping and online updates: Based on the posterior probability of the feature vector belonging to each window at the current time, select the window category corresponding to the highest probability, and map to obtain the tolerance parameter set {staying frequency threshold} corresponding to that category. Interruption tolerance time threshold Retention time retention coefficient Residence time decay coefficient The cluster centers and model parameters are updated using a weighted average method to adapt to the evolution of fish circadian rhythms.

8. The method for determining abnormal lingering behavior of fish in sturgeon farming ponds according to claim 1, characterized in that: Step S6 includes the following sub-steps: Step S601: Obtain the bottom distribution density, movement intensity and dwell time of the fish school, and obtain the score of abnormal dwelling state of the fish school based on the weighted calculation model; The scoring formula for abnormal stay status is as follows: ； in, The density of fish at the bottom of the school. Exercise intensity The normalized value, For normalized Duration of stay; Step S602: Compare the abnormal stay status score with a preset threshold; Step S603: When the abnormal stagnation status score exceeds the threshold, it is determined that the fish group is in an abnormal stagnation status.

Images