An underwater fish swarm monitoring and statistical system based on image fusion

Abstract

This invention relates to the field of underwater fish swarm monitoring technology and discloses an underwater fish swarm monitoring and statistical system based on image fusion. The system acquires underwater video streams and sonar reflection intensity data in different spectral bands through an underwater multi-source image acquisition module, and generates a spatiotemporally synchronized multimodal image data stream by aligning with timestamps. A fish swarm contour reconstruction module segments the fish swarm contour boundaries and fuses visible light textures and sonar geometric features to generate an underwater three-dimensional fish swarm distribution set. A dynamic trajectory mapping module tracks the centroid displacement, calculates the movement speed and directional offset angle, and correlates with water depth to generate a dynamic trajectory topology map. A behavior anomaly analysis module extracts environmental data based on trajectory mutation nodes, detects changes in aggregation density and directional dispersion to mark anomalous feature clusters. A population statistics output module integrates the data, classifies and statistically analyzes population distribution, quantity thresholds, and migration path overlap, and finally generates a statistical heatmap of fish swarm distribution.

Description

Technical Field

[0001] This invention relates to the field of underwater fish monitoring technology, specifically an underwater fish monitoring and statistical system based on image fusion. Background Technology

[0002] Underwater fish monitoring is a crucial component of fisheries resource management, aquatic ecosystem protection, and aquatic ecosystem research. Accurately understanding the distribution, quantity, movement patterns, and behavioral characteristics of fish schools provides critical information for sustainable fisheries development and maintaining aquatic ecological balance. However, the complex and unique underwater environment, where light propagation is easily affected by water scattering and absorption, presents numerous challenges to traditional visible light-based monitoring techniques. In turbid or deep water areas, visible light images often suffer from low contrast and blurred details, making it difficult to clearly capture the outlines and textures of fish schools, significantly limiting the accuracy of fish identification and statistics.

[0003] Single-sensor monitoring methods have significant limitations. Some monitoring systems rely solely on sonar technology, which, while able to penetrate turbid water to obtain the geometric outline and location information of targets, lacks textural detail, making it difficult to distinguish fish species and individual characteristics, and easily leading to misidentification of fish schools with other underwater objects. Furthermore, traditional monitoring systems often suffer from asynchronous data acquisition; the temporal discrepancy between visible light images and sonar data can cause errors in subsequent data fusion and analysis, affecting the accuracy of tracking fish dynamic trajectories.

[0004] In fish school contour reconstruction, traditional methods often rely on a single data source, making it difficult to balance contour integrity and detail richness. Visible light-based contour extraction is susceptible to water body interference, resulting in breaks or redundancy, while sonar-based contour extraction lacks fine structure due to resolution limitations, leading to poor 3D fish school distribution reconstruction results. During dynamic trajectory tracking, factors such as water flow disturbance and rapid fish movement can cause centroid identification errors. Traditional systems lack correlation analysis of environmental parameters such as water depth, causing discrepancies between the calculated motion speed and direction and the actual situation, making it difficult to accurately reflect the true movement state of the fish school.

[0005] In the analysis of abnormal behavior, traditional techniques often rely solely on changes in the fish's own movement parameters, neglecting the influence of environmental factors such as water turbulence intensity and temperature gradients on fish behavior, leading to a high false positive rate in abnormal behavior detection. Furthermore, in population statistics, traditional systems struggle to accurately distinguish between different fish species, and their statistics on distribution area, population thresholds, and migration path overlap are insufficient, resulting in statistical results that fail to meet the needs of refined management and research. These problems make it difficult for existing underwater fish monitoring systems to provide comprehensive and accurate fish population information in practical applications, hindering the in-depth development of fisheries resource management and ecological protection efforts. Summary of the Invention

[0006] The purpose of this invention is to provide an underwater fish school monitoring and statistics system based on image fusion to solve the problems mentioned in the background art.

[0007] To achieve the above objectives, the present invention provides an underwater fish swarm monitoring and statistics system based on image fusion, the system comprising:

[0008] The underwater multi-source image acquisition module is used to acquire underwater video streams in different spectral bands, simultaneously acquire sonar reflection intensity data, perform timestamp alignment processing on visible light images and sonar scan images, and generate spatiotemporally synchronized multimodal image data streams.

[0009] The fish school outline reconstruction module is used to extract the visible light image frame sequence and sonar intensity matrix from the multimodal image data stream, segment the fish school outline boundary based on the sonar echo intensity difference, and fuse the texture features of the visible light image and the sonar geometric outline features to generate an underwater three-dimensional fish school distribution set.

[0010] The dynamic trajectory mapping module is used to identify the centroid coordinates of each fish target in the underwater three-dimensional fish distribution set, continuously track the centroid displacement vector of adjacent time stamps, calculate the fish movement speed and direction offset angle, and generate a dynamic trajectory topology map by associating water depth data.

[0011] The abnormal behavior analysis module is used to extract the water flow turbulence intensity and temperature gradient data at the corresponding time points based on the abrupt movement rate nodes in the dynamic trajectory topology map, detect the changes in fish aggregation density and the dispersion of movement direction, and mark abnormal behavior feature clusters.

[0012] The population statistics output module is used to integrate the associated nodes of the abnormal behavior feature clusters and the dynamic trajectory topology map, classify and statistically analyze the distribution area, population size threshold and migration path overlap of different fish species, and generate a statistical heat map of fish population distribution.

[0013] Preferably, the underwater multi-source image acquisition module outputs infrared spectral image data, a sonar depth coding matrix, and a time synchronization identifier;

[0014] The output of the fish school outline reconstruction module includes a fused outline boundary coordinate set, texture-geometric feature matching score, and three-dimensional spatial distribution topology.

[0015] The dynamic trajectory mapping module outputs a centroid displacement vector sequence, a motion rate change curve, and a depth-direction correlation mapping table.

[0016] The output of the behavior anomaly analysis module includes a cluster of turbulence intensity anomaly markers, a density dispersion comparison coefficient, and a set of spatiotemporal coordinates of anomalies.

[0017] The population statistics output module outputs vector boundaries of species distribution areas, population density distribution matrix, and migration path overlay relationship diagram.

[0018] Preferably, the underwater multi-source image acquisition module includes:

[0019] The spectral image alignment submodule is used to receive underwater video streams from visible light and infrared camera devices, extract the exposure timestamp and spectral band identifier of each frame, align the multispectral image sequence through a timestamp deviation compensation algorithm, and generate a band-aligned image set.

[0020] The sonar data fusion submodule is used to obtain the depth reflection matrix of the sonar scanning device, map the spatial coordinates of the band-aligned image set to the sonar depth grid, calculate the overlapping coverage area of ​​the visible light pixel coordinates and the sonar grid, and output a spatiotemporally fused multimodal image data stream.

[0021] Preferably, the fish school outline reconstruction module includes:

[0022] The sonar contour segmentation submodule is used to parse the sonar intensity matrix in the multimodal image data stream, mark the potential fish school boundary according to the abrupt change point of the echo intensity gradient, extract the geometric center coordinates and area ratio parameters of the boundary closed region, and generate an initial fish school contour set.

[0023] The multi-source feature fusion submodule is used to compare the texture feature distribution of the initial fish school outline set with that of the visible light image, identify the scale texture region in the visible light image that overlaps with the sonar outline, perform weighted superposition calculation on the edge coordinates of the texture region and the sonar geometric outline, and reconstruct the underwater three-dimensional fish school distribution set.

[0024] Preferably, the dynamic trajectory mapping module includes:

[0025] The centroid displacement tracking submodule is used to extract the spatial coordinates of the centroid of the fish from the continuous time series of the underwater three-dimensional fish distribution set, calculate the Euclidean distance and depth change between adjacent time stamps, and generate a centroid displacement vector sequence.

[0026] The environmental association mapping submodule is used to obtain the real-time water flow velocity matrix collected by the water area sensor, calculate the difference between the motion direction angle and the water flow direction angle in the centroid displacement vector sequence, and construct a dynamic trajectory topology map by combining the temperature sensor data.

[0027] Preferably, the behavior anomaly analysis module includes:

[0028] The mutation node detection submodule is used to scan the coordinates of nodes whose motion rate exceeds a set threshold in the dynamic trajectory topology map, and extract the water flow turbulence intensity sampling value and temperature gradient change corresponding to the node.

[0029] The dispersion analysis submodule is used to calculate the standard deviation of the fish population density centered on the mutation node, the distribution dispersion coefficient of the fish movement direction angle within the statistical radius threshold range, and to label abnormal behavior feature clusters.

[0030] Preferably, the population statistics output module includes:

[0031] The species classification submodule is used to classify the outline feature labels of different fish species based on the texture feature difference values ​​of the underwater three-dimensional fish distribution set, and to mark the boundary coordinates of the spatial distribution area of ​​each type of fish.

[0032] The heatmap generation submodule is used to count the peak number of fish targets in each distribution area and the overlap rate of migration paths, map the number density to the chromaticity gradient value of the water space grid, and generate a statistical heatmap of fish population distribution.

[0033] Preferably, the system further includes:

[0034] The light source compensation optimization module is used to analyze the visible light image clarity score of the underwater multi-source image acquisition module. When the score is lower than a set threshold, the underwater supplementary lighting device is activated. The supplementary lighting intensity and angle parameters are adjusted according to the image grayscale mean distribution, and the optimized visible light image sequence is output to the fish school outline reconstruction module.

[0035] Preferably, the light source compensation optimization module includes:

[0036] The grayscale analysis submodule is used to calculate the mean and variance of grayscale values ​​in local areas of visible light image frames and to identify the spatial coordinate range of low-contrast regions.

[0037] The supplementary lighting control submodule is used to generate angle deflection commands for the supplementary lighting device based on the coordinate distribution of low-contrast areas, dynamically adjust the supplementary lighting intensity level in combination with water turbidity data, and output an optimized visible light image sequence.

[0038] Preferably, the system further includes:

[0039] The data correction module is used to receive the fish population distribution statistical heat map from the population statistics output module, extract the population density benchmark value from the historical monitoring data of the same period, calculate the deviation rate between the current distribution data and the benchmark value, perform spatial coordinate correction on the area where the deviation rate exceeds the threshold, and output the corrected fish population statistical distribution map.

[0040] Compared with the prior art, the beneficial effects of the present invention are:

[0041] This system provides a more comprehensive and accurate solution for underwater fish monitoring and statistics through multi-source data fusion and multi-module collaborative operation. The underwater multi-source image acquisition module integrates underwater video streams and sonar reflection intensity data from different spectral bands, achieving spatiotemporal synchronization through timestamp alignment, thus overcoming the limitations of a single sensor. Visible light images provide rich texture information, while sonar data has strong underwater penetration capabilities. Combining the two can compensate for their respective shortcomings in the underwater environment, providing more reliable basic data for subsequent analysis and avoiding information loss or bias caused by the deficiencies of a single data source.

[0042] The fish school contour reconstruction module segments contour boundaries based on sonar echo intensity differences, while simultaneously fusing texture features from visible light images with geometric contour features from sonar. The resulting underwater 3D fish school distribution set possesses both structural integrity and rich detail. Compared to traditional contour extraction methods using a single data source, this fusion processing effectively reduces the impact of water scattering and attenuation on image quality and compensates for the lack of texture information in sonar images. This results in a clearer and more accurate representation of the fish school's morphological distribution, laying a solid foundation for subsequent trajectory tracking and population analysis.

[0043] The dynamic trajectory mapping module identifies the centroid coordinates and continuously tracks the displacement vector, combining this with water depth data to generate a dynamic trajectory topology map, which more realistically reflects the movement state of the fish school. The calculation of movement speed and directional offset angle is no longer limited to changes in planar position but incorporates depth-dimensional information, making the movement trajectory of the fish school in three-dimensional underwater space more accurate. This reduces trajectory analysis errors caused by ignoring depth factors, and presents a more complete picture of the fish school's migration path and activity range.

[0044] The abnormal behavior analysis module correlates abrupt changes in fish movement rates with environmental data such as water flow turbulence intensity and temperature gradients. By detecting changes in aggregation density and the dispersion of movement direction, it identifies anomalous behavioral feature clusters, enabling multi-dimensional judgment of abnormal fish behavior. This analysis method, which incorporates environmental factors, overcomes the limitations of traditional methods that rely solely on the fish's own movement parameters. It can more accurately distinguish between natural movement changes and abnormal behavior, making anomaly labeling more consistent with actual ecological scenarios and avoiding misjudgments caused by relying on a single parameter.

[0045] The population statistics output module integrates the associated nodes of abnormal behavior feature clusters and dynamic trajectory topology maps, classifies and statistically analyzes the distribution area, population size threshold, and migration path overlap of different fish species, and generates heat maps, providing rich statistical information for fish population research and management. Through the integrated analysis of multi-dimensional data, it not only clearly presents the spatial distribution characteristics of different fish populations but also reflects the trends in population size changes and migration patterns, making the statistical results more valuable and meeting the diverse needs for fish population information in scenarios such as fisheries resource management and ecological protection. The various modules work synergistically, forming a complete technical chain from data collection, contour reconstruction, trajectory tracking, anomaly detection, and population statistics, comprehensively improving the accuracy and comprehensiveness of underwater fish population monitoring and statistics. Attached Figure Description

[0046] Figure 1 This is a schematic diagram illustrating the working principle of the underwater fish school monitoring and statistics system based on image fusion described in this invention.

[0047] Figure 2 Output a flowchart for each module;

[0048] Figure 3 This is a schematic diagram illustrating the working principle of an underwater multi-source image acquisition module.

[0049] Figure 4 This is a schematic diagram illustrating the working principle of the dynamic trajectory mapping module. Detailed Implementation

[0050] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0051] Please see Figure 1 This invention provides an underwater fish swarm monitoring and statistics system based on image fusion, the system comprising:

[0052] The system comprises an underwater multi-source image acquisition module, a fish school outline reconstruction module, a dynamic trajectory mapping module, a behavior anomaly analysis module, and a population statistics output module. The underwater multi-source image acquisition module acquires underwater video streams in different spectral bands and simultaneously acquires sonar reflection intensity data. By performing timestamp alignment processing on visible light images and sonar scan images, it generates a spatiotemporally synchronized multimodal image data stream. The fish school outline reconstruction module extracts the visible light image frame sequence and sonar intensity matrix from the multimodal image data stream. Based on the difference in sonar echo intensity, it segments the fish school outline boundary and fuses the texture features of the visible light image with the sonar geometric outline features to generate an underwater 3D fish school distribution set. The dynamic trajectory mapping module identifies the centroid coordinates of each fish target in the underwater 3D fish school distribution set, continuously tracks the centroid displacement vector of adjacent timestamps, calculates the fish school's movement speed and directional offset angle, and correlates it with water depth data to generate a dynamic trajectory topology map. The abnormal behavior analysis module extracts water flow turbulence intensity and temperature gradient data at corresponding time points based on the abrupt change nodes in the dynamic trajectory topology map, detects changes in fish population density and dispersion of movement direction, and labels abnormal behavior feature clusters. The population statistics output module integrates the abnormal behavior feature clusters with the associated nodes in the dynamic trajectory topology map, classifies and statistically analyzes the distribution area, population size threshold, and migration path overlap of different fish species, and generates a statistical heatmap of fish population distribution.

[0053] Example 1: See Figure 2 The infrared spectral image data output by the underwater multi-source image acquisition module is generated by infrared camera equipment deployed in the monitoring area. The infrared camera equipment operates in the wavelength range of 750 nm to 1450 nm, a band whose water penetration capability is superior to that of the visible light spectrum. Each infrared image frame is accompanied by a precise timestamp recorded by the device's built-in clock, accurate to the millisecond level. The infrared video stream is continuously captured at a sampling rate of 25 frames per second, with a spatial resolution of 1920 × 1080 pixels per frame. The infrared spectral image data is stored in a grayscale encoding format, with grayscale values ​​mapping the surface temperature distribution of the object, and high-temperature areas exhibiting high brightness characteristics. During the acquisition process, the device synchronously records spectral band identification parameters, including the spectral center wavelength value and bandwidth range, which are embedded in the image file header in the form of metadata.

[0054] The sonar depth coding matrix is ​​generated by a multibeam sonar scanning device. The sonar device operates at a frequency of 200 kHz, with a scanning angle covering a 120-degree sector and a scan line spacing of 0.5 degrees. Each scan generates a depth reflection matrix with dimensions of 240 × 500 (number of scan lines × number of distance cells). Each element in the matrix contains two data values: a depth value in meters and a reflection intensity value of a 16-bit integer (0-65535). Spatial coordinates are recorded in a polar coordinate system and mapped to a three-dimensional Cartesian coordinate system using a transformation algorithm, forming an underwater spatial grid coordinate array. The spatial resolution of this grid is 0.1 m × 0.1 m × 0.1 m, covering an effective range of 5 m × 5 m × 3 m within the monitoring area.

[0055] The generation of the time synchronization identifier relies on a high-precision time synchronization protocol. Visible light cameras, infrared devices, and sonar scanners are each connected to a GPS clock synchronization module. Each device receives a UTC time signal once per second and calibrates its internal clock. The timestamp alignment process involves two key operations: first, extracting the deviation between the timestamp of each device and the UTC reference time; second, applying a linear interpolation algorithm to compensate for the time difference in data from asynchronous sampling points. In the aligned data stream, each frame of visible light image, infrared image, and sonar scan data carries a unified time identifier. This identifier is a 32-bit integer, with the first 16 bits representing the date and the last 16 bits representing the number of milliseconds in that day.

[0056] The fused contour boundary coordinate set output by the fish swarm contour reconstruction module comes from the multimodal data processing. The sonar intensity matrix is ​​processed by an edge detection algorithm, and the Sobel operator convolution kernel scans the region of reflection intensity gradient change. When the intensity difference between adjacent grid points exceeds 1500, it is marked as a contour boundary point. Boundary points are automatically connected to form closed polygon regions, and the vertex coordinates of the polygons are stored as a three-dimensional floating-point array. The texture-geometric feature matching score is calculated using a feature comparison method: the texture feature vector (including the contrast of the gray-level co-occurrence matrix and correlation parameters) of the corresponding contour region in the visible light image is extracted and similarity is calculated with the geometric features of the sonar contour (including area, perimeter, and curvature). The similarity value is normalized to a 0-1 score. The three-dimensional spatial distribution topology is represented by an octree data structure, and the side length of the cube region occupied by each fish swarm target fluctuates within 1.2 times its maximum contour diameter. The topology records the spatial adjacency relationship and hierarchical structure between the cubes.

[0057] The centroid displacement vector sequence generation process of the dynamic trajectory mapping module is as follows: Spatial coordinate data of 20 consecutive time points are extracted from the 3D fish swarm distribution set, with an adjacent time interval of 200 milliseconds. The 3D spatial centroid coordinates of each fish swarm target are calculated using a centroid positioning algorithm, with a coordinate accuracy of 0.01 meters. The displacement vector calculation uses the Euclidean distance formula, and the result includes the 3D displacement distance and horizontal direction angle. The motion rate change curve is generated by the first derivative of the displacement time series, with a sampling frequency of 50 Hz. The curve data is stored as a time-rate key-value pair list. The depth-direction association mapping table adopts a multi-dimensional matrix structure. The first column of the matrix is ​​the timestamp index, the second column is the depth value, and the third to fifth columns record the values ​​of the horizontal direction angle, vertical angle, and velocity, respectively. The number of rows in the matrix corresponds to the number of time points of trajectory sampling.

[0058] The turbulence intensity anomaly marker cluster detection mechanism of the behavior anomaly analysis module is as follows: A water flow turbulence sensor collects turbulence intensity values ​​at a frequency of 50 Hz, with the unit being cm / s². Turbulence data is matched with time nodes in the dynamic trajectory topology map. When the turbulence value at a spatial node exceeds the critical value of 35 cm / s², the node coordinates are marked as an anomaly point. The density dispersion comparison coefficient calculation is divided into three stages: first, a 2-meter radius area is defined, and the standard deviation of the fish population distribution is statistically analyzed; second, the variance of the movement direction angles of all individuals within the area is calculated; finally, the ratio of the standard deviation to the variance is obtained as the output data. The spatiotemporal coordinate set of anomaly behavior is recorded using a four-dimensional data packet structure. The data within the packet includes fields such as date and timestamp, three-dimensional spatial coordinates, and anomaly type code.

[0059] The species distribution area vector boundary of the population statistics output module is generated using the Delaunay triangulation algorithm. Based on the spatial distribution coordinate point set of fish schools, the algorithm automatically constructs an irregular triangular network and selects the boundary of the density core region as the vector contour. The contour data is stored in SVG graphic format as a vertex coordinate sequence. The population density distribution matrix is ​​constructed using a spatial grid partitioning method, dividing the water body into 0.5m × 0.5m × 0.5m cubic cells, and the number of fish in each cell is counted to generate a three-dimensional density matrix. The calculation of the migration path overlay relationship map focuses on path correlation analysis: extracting the center lines of the movement trajectories of different fish species, and calculating the minimum distance between each trajectory; when the distance is less than 1 meter for 30 consecutive seconds, it is defined as path overlap. The overlay relationship map uses different color bands to mark the overlapping areas of the migration paths of various fish species.

[0060] The format conversion rules for the output data of the above modules are as follows: infrared spectral images are encapsulated in TIFF format to encapsulate time synchronization identifier metadata; sonar depth coding matrices are stored in HDF5 file format for 3D mesh data; fused contour boundary coordinate sets are recorded in GeoJSON format for polygon data; 3D spatial distribution topology relationships are efficiently stored and retrieved using a custom binary data structure; centroid displacement vector sequences are stored in a time-series database indexed by timestamps; and anomalous behavior spatiotemporal coordinate sets are transmitted across platforms using the ProtocolBuffers data serialization standard. All output data is transmitted between modules in real time via message queues, and each processing node deploys a data format parsing plugin to extract information. The anomaly verification mechanism during data processing includes timestamp continuity checks, coordinate boundary validity verification, and data range threshold alarms.

[0061] Example 2: See Figure 3 The underwater multi-source image acquisition module's spectral image alignment submodule receives two independent video streams from a visible light camera and an infrared camera. The visible light camera uses a 1-inch CMOS sensor with 20 million effective pixels, a frame rate of 30fps, and operates in the 400-700 nm visible spectrum. The infrared camera is equipped with a cooled mid-wave infrared detector with a response band of 3-5 micrometers, a resolution of 640×512 pixels, and a frame rate of 25fps. Both video streams are transmitted to the processing unit via a gigabit Ethernet interface. Each frame carries a timestamp generated by the device's internal clock. The timestamp includes year, month, day, hour, minute, second, and millisecond information, with an accuracy of ±1 millisecond. Spectral band identification information is embedded in the image file's metadata area; the visible light image is labeled "VIS," and the infrared image is labeled "IR-MW," along with the center wavelength value and full width at half maximum (FWHM) parameter.

[0062] The timestamp offset compensation algorithm performs multispectral image sequence alignment. The algorithm first parses the timestamps of the two video streams and calculates the clock offset between the devices. Clock offset compensation uses a sliding window comparison method with a window size of 5 seconds. Keyframes from the visible light video stream are selected as references within the window, and corresponding frames are found in the infrared video stream through feature point matching. Feature points are extracted using an accelerated robust feature algorithm, and the matching process employs a nearest neighbor search method. For successfully matched frame pairs, the time offset value is recorded, and the clock drift curve between the devices is fitted using the least squares method. The final alignment operation uses cubic spline interpolation to resample the infrared image sequence, achieving frame-level synchronization between the two video streams. The synchronization error is controlled within ±2 milliseconds. The generated band-aligned image set is stored on a solid-state drive array, with each frame accompanied by a unified timestamp and spectral identifier.

[0063] The sonar data fusion submodule processes the depth reflection matrix acquired by the multi-beam sonar scanning device. The sonar device operates at a frequency of 455 kHz, with a beam opening angle of 1° × 1°, a maximum detection range of 100 meters, and a range resolution of 1 centimeter. Each scan generates reflection data for 240 beams, each beam containing 500 range cells. The depth reflection matrix is ​​stored in polar coordinates, with each data point containing distance, reflection intensity, and beam angle information. The coordinate transformation process first establishes a three-dimensional Cartesian coordinate system, with the origin located at the center of the sonar transducer. Using a polar-to-rectangular coordinate transformation formula, each reflection point is mapped to a three-dimensional spatial grid. The grid resolution is set to 0.05 m × 0.05 m × 0.05 m, covering a 10 m × 10 m × 5 m monitoring area.

[0064] The registration process between visible light pixel coordinates and the sonar grid employs a feature matching method. SIFT feature points are extracted from the band-aligned image set, while strong reflectivity feature regions are extracted from the sonar intensity image. The projection transformation matrix from the image plane to the sonar grid is calculated using the RANSAC algorithm. The registration accuracy reaches within 2 pixel error range. The overlapping region calculation uses spatial hashing to establish a bidirectional mapping relationship between image pixels and the sonar grid. The fusion process weights and superimposes visible light texture and sonar reflectivity, with the weighting coefficients dynamically adjusted according to water depth and turbidity. The final generated multimodal image data stream adopts a hierarchical storage structure: the bottom layer is the sonar depth grid, the middle layer is the registered visible light image, and the top layer is the infrared temperature distribution data. The data stream is transmitted to the fish school contour reconstruction module via a 10G fiber optic network.

[0065] The sonar contour segmentation submodule of the fish swarm contour reconstruction module processes sonar intensity information from the multimodal data stream. Median filtering is used to eliminate noise in the intensity matrix preprocessing, with a window size of 3×3×3. Edge detection uses the 3D Sobel operator, calculating gradients in the X, Y, and Z directions. The gradient threshold is set to 15% of the reflection intensity value; points exceeding the threshold are marked as boundary candidate points. The boundary connection algorithm employs a 3D region growing method, selecting the location of the maximum gradient as the seed point, with growth conditions including gradient direction and intensity continuity. Closed region detection is achieved by calculating the topological relationships of boundary points, using the MarchingCubes algorithm to construct isosurfaces. The geometric properties of each closed region are calculated, including volume, surface area, principal axis direction, and density. The initial fish swarm contour set is stored in 3D point cloud format, with each contour accompanied by a 128-dimensional feature vector describing its geometric characteristics.

[0066] The multi-source feature fusion submodule performs the matching process between visible light texture and sonar contour. It extracts the target fish area from the registered visible light image and analyzes texture features using a local binary mode algorithm. The texture feature vector contains three components: uniformity, contrast, and correlation. Geometric feature matching employs a nearest neighbor search method to find the optimal correspondence between visible light texture and sonar contour in the feature space. Successfully matched feature pairs enter the weighted fusion stage, with fusion weights dynamically calculated based on feature matching degree and sensor confidence. Texture feature weights increase in clear water, while sonar geometric feature weights increase in turbid water. The fusion process generates an enhanced 3D fish school contour, with the contour surface simultaneously containing geometric shape information and texture details. The reconstructed underwater 3D fish school distribution set is organized using an octree structure, with each node storing spatial location, contour attributes, and texture features. The distribution set is transmitted to the dynamic trajectory mapping module via a high-speed data interface.

[0067] Key technical parameters involved in system implementation include: time synchronization accuracy ±2 milliseconds, spatial registration error less than 3 centimeters, feature matching success rate above 85%, and contour reconstruction completeness above 90%. The data processing pipeline adopts a parallel computing architecture, with visible light, infrared, and sonar data processed by dedicated processing units, and data synchronization performed during the fusion phase. The entire system runs on an embedded Linux platform, with core algorithms implemented in C++ and optimized for the NEON instruction set. Real-time performance meets the requirement of processing 25 frames of multimodal data per second, with latency controlled within 200 milliseconds. Data storage uses a RAID5 disk array to ensure the integrity and reliability of the acquired data. The system network architecture adopts a layered design, with the sensor layer connected via industrial Ethernet and the processing layer interconnected via InfiniBand network to meet the needs of large data volume transmission.

[0068] Example 3: See Figure 4 The centroid displacement tracking submodule of the dynamic trajectory mapping module processes 3D fish swarm distribution data from the fish swarm contour reconstruction module. The input data is organized using an octree structure, with each leaf node storing complete descriptive information for a fish swarm target. The centroid coordinate calculation process first traverses the octree structure to extract the coordinates of all surface vertices for each fish swarm target. The vertex coordinate set undergoes denoising filtering to remove outliers that significantly deviate from the main distribution. The filtering algorithm uses statistical outlier detection to remove data points that deviate from the mean by more than three standard deviations. The centroid position is determined using a weighted average method, with the weighting factor related to the vertex density distribution. The calculation process is iterative until the change in centroid position between two consecutive iterations is less than 0.01 meters. The final centroid coordinate accuracy reaches the centimeter level, meeting the accuracy requirements for trajectory tracking.

[0069] The displacement calculation for adjacent timestamps employs a sliding time window method. The system sets a fixed time window of 200 milliseconds, matching fish targets between consecutive frames within this window. The matching criteria comprehensively consider spatial distance, volume similarity, and texture feature correlation. The spatial distance threshold is set to 1.5 times the average diameter of the fish swarm; volume similarity is required to vary by no more than 20%; and cosine similarity is used to measure texture feature correlation. Successfully matched fish targets proceed to the displacement calculation process, using the Euclidean distance formula as follows:

[0070]

[0071] in, Indicates displacement distance. and Representing adjacent times and The centroid coordinates at any given time. The depth change is directly taken as the difference in the Z coordinates. The orientation angle is calculated using a spherical coordinate system transformation; the horizontal orientation angle... Vertical direction angle The generated centroid displacement vector sequence is stored as a time series database, with each record containing a timestamp, displacement, depth change, and orientation angle.

[0072] The environmental correlation mapping submodule integrates multi-source environmental sensor data. A water flow velocity sensor array acquires three-dimensional flow velocity data at a frequency of 10Hz, with a measurement range of 0-3m / s and an accuracy of ±0.02m / s. The temperature sensor uses platinum resistance elements, with a measurement range of 0-30℃ and a resolution of 0.01℃. Sensor data is transmitted to the processing unit via a fieldbus network, achieving a time synchronization accuracy of 5 milliseconds. The water flow direction angle is calculated using a vector synthesis method, projecting the three-dimensional flow velocity components onto a horizontal plane to determine the direction of the synthesized vector. The temperature gradient is calculated using the spatial difference method, determining the rate of temperature change within a 1-meter cubic grid.

[0073] The difference between the motion direction angle and the water flow direction angle is calculated using an angular distance metric. First, the two angle values ​​are normalized to the interval [0, 2π], and then the minimum circumferential distance is calculated. Difference index Defined as:

[0074]

[0075] in, Indicates the angle of water flow direction. This represents the directional angle of the fish school's movement. A smaller difference value indicates a higher consistency between the fish school's movement and the water flow direction. Temperature influence factors were obtained by establishing a temperature-velocity regression model, with model parameters updated online to adapt to environmental changes. The dynamic trajectory topology map was constructed using a graph data structure, where nodes represent the centroid positions of the fish school, and edges store the correlation between movement attributes and environmental parameters. The spatiotemporal resolution of the graph was set to a 0.1-meter spatial grid and a 200-millisecond time interval.

[0076] The mutation node detection submodule of the behavior anomaly analysis module scans the dynamic trajectory topology to identify abnormal changes in motion state. The velocity threshold is dynamically adjusted based on historical data, and an adaptive algorithm is used to calculate the normal motion range under current environmental conditions. The threshold calculation window is set to a 30-second sliding window, covering 150 consecutive sampling points. Wavelet analysis is used to sample water flow turbulence intensity, extracting energy features in the 0.1-1Hz frequency band as a turbulence index. Temperature gradient changes are calculated using the five-point central difference method, reflecting the spatiotemporal variation characteristics of the local temperature field.

[0077] The dispersion analysis submodule processes the spatial statistical characteristics of fish school aggregation. Aggregation density is calculated using kernel density estimation, with the bandwidth of the Gaussian kernel function set to twice the average diameter of the fish school. The density standard deviation is calculated within a spherical analysis region, the radius of which is proportional to the size of the fish school. The directional dispersion coefficient is obtained through vector statistical analysis, calculating the spherical variance of the movement direction angles of all individuals. The labeling process for anomalous behavior feature clusters employs a density clustering algorithm to identify high-density anomalous regions in the spatiotemporal-parametric multidimensional space. The clustering results are stored as a timestamped spatial envelope, recording the start time, duration, and spatial extent of the anomalous behavior.

[0078] In the system implementation, the dynamic trajectory mapping module adopts a distributed computing architecture. The centroid tracking task is distributed across multiple computing nodes for parallel processing, with each node responsible for the fish swarm target in a specific spatial region. Nodes exchange boundary target information via a high-speed network to maintain trajectory continuity. Environmental correlation analysis runs on a dedicated signal processing unit, enabling real-time fusion processing of sensor data. Anomaly detection algorithms are deployed on a GPU-accelerated platform, utilizing parallel computing capabilities to process large-scale topology map data. Data storage employs a time-series database cluster, supporting high-throughput trajectory data writing and querying.

[0079] The network communication adopts a layered architecture. The sensor layer uses the real-time Ethernet protocol to ensure the timeliness of data acquisition. The processing layer adopts a publish / subscribe messaging pattern, with modules exchanging data through message queues. The display layer uses the WebSocket protocol to achieve real-time visualization of monitoring results. System time synchronization adopts the IEEE 1588 precise time protocol, with all nodes maintaining microsecond-level clock synchronization. The data processing pipeline is designed as a stateful service, maintaining the context information of trajectory tracking to ensure the continuity of motion analysis.

[0080] The hardware platform is optimized for underwater environments. The main processing unit uses a ruggedized industrial computer with IP68 protection. The signal acquisition module features electromagnetic compatibility design to resist underwater electrical interference. All equipment has passed pressure chamber testing to ensure normal operation at specified depths. The power system employs a redundant design, with automatic switching between main power and backup batteries. The system deployment takes into account underwater topology, optimizing sensor placement to reduce acoustic multipath interference and optical obstruction effects.

[0081] The software architecture employs a modular design. Core algorithms are encapsulated as independent services, interacting through well-defined interfaces. Configuration management utilizes version control, supporting online parameter adjustment and hot algorithm updates. Error handling mechanisms include data verification, anomaly capture, and automatic recovery. The logging system records detailed operational status, supporting post-event analysis and fault diagnosis. The user interface offers multiple view modes, allowing interactive exploration of fish movement characteristics and environmental correlation patterns. System maintenance tools support remote monitoring and diagnostics, reducing the need for underwater operations.

[0082] Example 4: This example involves the collaborative operation of the population statistics output module and the light source compensation optimization module. The implementation process is detailed below with a specific example. The example is based on a scenario of monitoring tuna schools in coastal waters: the system is deployed in a 10-meter deep water area, covering a 500m × 500m area. The underwater multi-source image acquisition module has generated a spatiotemporally synchronized multimodal image data stream, and the fish school contour reconstruction module outputs an underwater 3D fish school distribution set. This distribution set contains the 3D coordinates and texture feature data of three main species: tuna, mackerel, and sardines. The species classification submodule begins processing this input data. The submodule classifies fish schools by texture feature difference values. Texture features are extracted from the tuna scale patterns in visible light images, including contrast values ​​(range 0.1-0.9) and correlation values ​​(range 0.8-1.0) in the gray-level co-occurrence matrix. The difference value calculation uses a feature vector distance metric: for each fish school target's texture vector (dimension 128), the optimal segmentation point is found in the feature space using a k-means clustering algorithm. Ultimately, three contour feature labels were identified: "TSK" (tuna), "MAC" (mackerel), and "SAR" (sardine). Next, the submodule labeled the spatial distribution boundary coordinates of each fish school. The boundary calculation employed the Alpha shape algorithm: extracting the 3D point cloud data for each label, setting the alpha radius to 5 meters, and forming a closed polygonal contour. Boundary coordinates were stored in vector format, with each set containing 100-200 vertex coordinates and an accuracy of 0.01 meters. For example, the boundary coordinate set for a tuna school was recorded as [(x1,y1,z1),(x2,y2,z2),...,(xn,yn,zn)], representing the extent of its core habitat.

[0083] The heatmap generation submodule receives the classified data and performs statistical operations. First, it counts the peak number of fish in each distribution area: the water area is divided into 0.5m × 0.5m × 0.5m grid cells, all cells are traversed, and the number of fish belonging to the same tag within each cell is counted. The peak number is the maximum value in the cell; for example, 85 tuna are recorded at the (100m, 150m) location in the monitoring area. Migration path overlap rate is calculated using a path intersection algorithm: the centerline of the movement trajectory of each fish species is extracted (provided by the dynamic trajectory mapping module), and the spatial and temporal overlap of different species' trajectories is checked. Overlap determination is based on the minimum Euclidean distance threshold between two trajectories (set to 1 meter) and temporal consistency (continuous overlap exceeding 30 seconds). For example, the migration paths of tuna and mackerel show a 45% overlap rate in the (200m-300m) water area. Subsequently, the submodule maps the number density to the chromaticity gradient values ​​of the water area's spatial grid. The mapping process uses the HSV color model: density values ​​are normalized to the range of 0-1, with 0 corresponding to blue (low density) and 1 corresponding to red (high density). The density value is calculated by dividing the average number of fish in each grid cell by the maximum cell capacity. The grid data is stored in a matrix structure. Table 1 shows a simplified fragment of the tuna population density distribution (header: grid cell ID, X coordinate, Y coordinate, depth layer, number of fish, color code). This data is used to generate the final statistical heatmap of the fish population distribution, which is in RGB image format, with pixel density representing fish abundance.

[0084] Table 1: Simplified distribution of tuna population density

[0085] Grid cell ID X-coordinate (m) Y coordinate (m) Depth layer (m) fish population Color coding GC-001 100 150 5 85 #FF0000 GC-002 102 152 5 60 #FF4500 GC-003 105 155 5 30 #00BFFF GC-004 110 160 10 10 #0000FF

[0086] The light source compensation and optimization module operates in parallel, analyzing the sharpness score of visible light images to improve input quality. The sharpness score is based on image processing algorithms: Laplacian edge detection is performed on each frame of the visible light image, and the variance of pixel gradient values ​​is calculated. The score ranges from 0 to 100, with a threshold of 60 (based on a preset standard for the water area). During monitoring, when the module detects that the image score in a densely populated tuna area drops to 55, the underwater supplemental lighting device is activated. The device is an LED array with adjustable intensity and angle. The grayscale analysis submodule calculates the mean and variance of grayscale values ​​in local image regions: the image is divided into 8×8 pixel blocks, and the mean and variance of each block are analyzed. For example, a low-contrast region (grayscale variance < 10, threshold set to 15) is identified in the (x=200m, y=250m) region. This coordinate is mapped to physical space using a coordinate transformation algorithm. Based on this, the supplemental lighting control submodule generates angle deflection commands for the supplemental lighting device: outputting servo control signals to adjust the device's pointing angle to 30 degrees azimuth and -10 degrees pitch. Intensity adjustment is combined with turbidity data (in NTU) provided by the water quality sensor: a turbidity value of 20 corresponds to intensity level 5 (the highest level). The optimized output sequence is transmitted to the fish school outline reconstruction module via a fiber optic interface.

[0087] The entire processing flow example integrates the above operations: It begins with a monitoring task where an underwater 3D fish distribution set is input into the species classification submodule. Tuna schools are identified and their boundaries are marked; the heat map generation submodule processes the grid data, statistically analyzes the density as shown in the table, and maps colors; simultaneously, the visible light image displays low resolution in deep water; the light source optimization module activates supplementary lighting (angle deflection 30 degrees, intensity level 5), outputting an optimized image. Finally, a statistical heat map of fish population distribution is generated, showing high-density areas of tuna in shallow water. The system adopts a modular interface design, with data processing latency controlled within 200 milliseconds, and data is stored in a distributed database. During execution, false data claims need to be avoided, and all operations are implemented based on algorithm and sensor input.

[0088] Example 5: The grayscale analysis submodule of the light source compensation optimization module begins processing the visible light image sequence from the underwater multi-source image acquisition module. The input image resolution is 4096×2160 pixels, with a color depth of 16 bits. The submodule performs regionalization processing, dividing each frame of the image into 512 uniform grid regions, each containing a 256×256 pixel block. Grayscale statistical analysis is performed on each pixel block: the average grayscale value of all pixels in the region is calculated, and the standard deviation of the grayscale value is also calculated. The analysis window moves across the image in a sliding manner with a step size of 128 pixels, ensuring that the region coverage has 50% overlap. The grayscale mean and variance calculation adopts a full-frame traversal algorithm, recording the statistical results of each grid region. Low-contrast regions are identified based on a preset variance threshold: when the standard deviation of a region is lower than 15 (range 0-255), it is determined to be a low-contrast region. The system records the coordinates of the bounding rectangle of all low-contrast regions, with the origin (0,0) at the top left corner of the image and the coordinates (4095, 2159) at the bottom right corner. The rectangular region is marked as a quadruple (starting X coordinate, starting Y coordinate, width, height) and stored in a buffer queue.

[0089] The supplemental lighting control submodule receives coordinate data of the low-contrast area and real-time water quality sensor information. The water quality sensor collects turbidity data at a frequency of 5Hz, in NTU units, with a measurement range of 0-100 NTU. A turbidity-intensity mapping table is pre-configured: turbidity values ​​of 0-10 NTU correspond to intensity level 1; for every 10 NTU increase in turbidity, the intensity level increases by one level, up to a maximum of level 10. Angle deflection commands are generated based on a coordinate transformation algorithm: first, the image coordinate system is converted to the device coordinate system, and the azimuth and pitch angles of the center point of the low-contrast area relative to the supplemental lighting position are calculated. For example, if the center point of a low-contrast area is at image coordinates (1200, 800), the converted azimuth angle is 35 degrees and the pitch angle is -15 degrees. The supplemental lighting device is a multi-array LED system with a mechanical rotation capability of ±180 degrees horizontally and ±45 degrees vertically. Control commands include angle parameters, duration, and intensity level, in binary control word format. After parsing the commands, the device drive circuit drives the rotation mechanism via a stepper motor, completing the angle adjustment within 500 milliseconds. Intensity adjustment is achieved through PWM pulse width modulation, with a duty cycle of 0-100% corresponding to a brightness of 0-20000 lumens. The optimized image sequence begins output 300 milliseconds after the supplementary lighting takes effect, and the frame sequence embeds supplementary lighting parameter metadata.

[0090] The data correction module processes the heatmap data generated by the population statistics output module. The input heatmap uses a three-dimensional grid structure with a grid size of 0.5m × 0.5m × 0.5m, and each grid cell stores the fish population density value. The historical database stores monitoring data from the same period over the past three years, with a data structure consistent with the current heatmap. The baseline extraction process performs spatiotemporal alignment: historical datasets with the same geographical coordinates and the same month and date are selected, and the average density value of each grid cell is calculated as the baseline. The deviation rate is calculated using fractional arithmetic: the current density value is subtracted from the baseline value, and then divided by the baseline value; the result is converted into a percentage. For example, if the current density of a grid is 85 fish and the historical baseline is 80 fish, the deviation rate is 6.25%. The system sets a deviation threshold of 20%. When the deviation rate of 10 consecutive grids in a certain area exceeds the threshold, spatial correction for that area is initiated.

[0091] Spatial coordinate correction employs an adaptive filtering algorithm. A three-dimensional buffer zone with a radius of 2 meters is established for the out-of-range area. The density value sequence of all grids within the buffer is taken, and a moving average filter with a window size of 3×3×3 is applied. The correction process iterates three times, adjusting the weighting coefficients in each iteration. The final generated new density values ​​are written into the correction dataset. Coordinate correction also includes position offset compensation: a random displacement of 0.1-0.3 meters is applied to the center point coordinates of the fish swarm target, with the displacement direction determined based on the water flow velocity vector field. After correction, the data undergoes a verification mechanism to confirm the continuity of the density distribution and eliminate anomalous abrupt changes. The output module generates the final statistical distribution map of the fish swarm, in a format compatible with GIS geographic information systems, including a vector boundary layer, a density raster layer, and a metadata attribute table.

[0092] During the closed-loop process, the data stream is strictly aligned to timestamps. The light source compensation operation performs an environmental assessment every 2 seconds, and the data correction process is automatically triggered every hour. All parameter adjustments are logged, with log entries including timestamps, operation types, input parameters, and output results. In the deep-sea monitoring case, the system processed a 24-hour dataset from a 600×600 meter sea area: grayscale analysis processing at a frame rate of 30fps, identifying an average of 28 low-contrast areas per frame; the maximum angle adjustment of the supplementary lighting equipment was 60 degrees; the data correction module processed a total of 172,800 heat map grids, correcting 320 areas exceeding the standard. No effect evaluation descriptions are provided during the processing; all operations are based on the algorithm's preset procedures and sensor measurement parameters. The system components interact with each other via a high-speed data bus, exchanging control signals and data packets.

[0093] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0094] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An underwater fish school monitoring and counting system based on image fusion, characterized in that, The system comprises: an underwater multi-source image acquisition module, configured to acquire underwater video streams of different spectral bands, synchronously acquire sonar reflection intensity data, perform timestamp alignment processing on visible light images and sonar scanning images, and generate a spatio-temporally synchronized multi-modal image data stream; a fish school contour reconstruction module, configured to extract visible light image frame sequences and sonar intensity matrices from the multi-modal image data stream, segment fish school contour boundaries based on sonar echo intensity differences, fuse texture features of visible light images and geometric contour features of sonar, and generate an underwater three-dimensional fish school distribution set; a dynamic trajectory mapping module, configured to identify centroid coordinates of each fish school target in the underwater three-dimensional fish school distribution set, continuously track centroid displacement vectors of adjacent timestamps, calculate fish school movement rates and direction deviation angles, and generate a dynamic trajectory topology graph in association with water depth data; a behavior anomaly analysis module, configured to extract water flow turbulence intensity and temperature gradient data at corresponding time points according to movement rate mutation nodes in the dynamic trajectory topology graph, detect fish school aggregation density change values and movement direction dispersion degrees, and mark abnormal behavior feature clusters; a population statistics output module, configured to integrate associated nodes of the abnormal behavior feature clusters and the dynamic trajectory topology graph, classify and statistically analyze distribution area, population quantity threshold and migration path coincidence degree of different fish school species, and generate a fish school quantity distribution statistical heat map; the dynamic trajectory mapping module comprises: a centroid displacement tracking sub-module, configured to extract fish school centroid spatial coordinates of continuous time sequences from the underwater three-dimensional fish school distribution set, calculate coordinate Euclidean distances and depth changes of adjacent timestamps, and generate a centroid displacement vector sequence; an environment correlation mapping sub-module, configured to acquire real-time water flow velocity matrices collected by water area sensors, calculate differences between movement direction angles and water flow direction angles in the centroid displacement vector sequence, and construct a dynamic trajectory topology graph in combination with temperature sensor data.

2. The underwater fish school monitoring and statistical system based on image fusion according to claim 1, wherein: the underwater multi-source image acquisition module outputs infrared spectral image data, sonar depth encoding matrices and time synchronization identifiers; the fish school contour reconstruction module outputs fused contour boundary coordinate sets, texture-geometry feature matching degree scores and three-dimensional spatial distribution topology relationships; the dynamic trajectory mapping module outputs centroid displacement vector sequences, movement rate change curves and depth-direction correlation mapping tables; the behavior anomaly analysis module outputs turbulence intensity anomaly marker clusters, density dispersion contrast coefficients and abnormal behavior spatio-temporal coordinate sets; the population statistics output module outputs species distribution area vector boundaries, population quantity density distribution matrices and migration path superposition relationship graphs. 3.The image fusion-based underwater fish school monitoring and counting system according to claim 1, wherein, the underwater multi-source image acquisition module comprises: a spectral image alignment sub-module, configured to receive underwater video streams of visible light and infrared cameras, extract exposure timestamps and spectral band identifiers of each frame of image, align multi-spectral image sequences through a timestamp deviation compensation algorithm, and generate a band-aligned image set. The sonar data fusion submodule is configured to acquire a depth reflection matrix of a sonar scanning device, map spatial coordinates in the band-aligned image set to a sonar depth grid, calculate an overlapping coverage area of the visible light pixel coordinates and the sonar grid, and output a spatio-temporal fused multi-modal image data stream.

4. The image fusion-based underwater fish school monitoring and counting system according to claim 3, wherein, The fish school profile reconstruction module includes: The sonar profile segmentation submodule is configured to analyze a sonar intensity matrix in the multi-modal image data stream, mark potential fish school boundaries according to echo intensity gradient abrupt change points, extract geometric center coordinates and area proportion parameters of a boundary closed region, and generate an initial fish school profile set. The multi-source feature fusion submodule is configured to compare the initial fish school profile set with texture feature distribution of the visible light image, identify a scale texture region in the visible light image that overlaps with the sonar profile, perform weighted superposition calculation on texture region edge coordinates and the sonar geometric profile, and reconstruct an underwater three-dimensional fish school distribution set. 5.The image fusion-based underwater fish school monitoring and counting system according to claim 1, wherein, The behavior anomaly analysis module includes: The abrupt node detection submodule is configured to scan node coordinates with a motion rate exceeding a set threshold in the dynamic trajectory topology graph, extract water flow turbulence intensity sampling values and temperature gradient change amounts corresponding to the node, and calculate a fish school aggregation density standard deviation centered on the abrupt node. The dispersion analysis submodule is configured to calculate a distribution dispersion coefficient of fish school motion direction angles within a radius threshold range, and mark an abnormal behavior feature cluster. 6.The image fusion-based underwater fish school monitoring and counting system according to claim 1, wherein, The population statistical output module includes: The species classification submodule is configured to divide profile feature labels of different fish school species according to texture feature difference values in the underwater three-dimensional fish school distribution set, and label spatial distribution region boundary coordinates of each fish school. The thermal map generation submodule is configured to calculate a number peak value of fish school targets in each distribution region and a migration path overlap rate, map the number density to a chroma gradient value of a water space grid, and generate a fish school number distribution statistical thermal map. 7.The image fusion-based underwater fish school monitoring and counting system according to claim 1, wherein, The system further includes: The light source compensation optimization module is configured to analyze a visible light image definition score of the underwater multi-source image acquisition module, activate an underwater light compensation device when the score is lower than a set threshold, adjust light compensation intensity and angle parameters according to image gray mean value distribution, and output an optimized visible light image sequence to the fish school profile reconstruction module. 8.The image fusion-based underwater fish school monitoring and counting system according to claim 7, wherein, The light source compensation optimization module includes: The gray analysis submodule is configured to calculate a local region gray mean value variance of a visible light image frame, and identify spatial coordinate ranges of low contrast regions. The light compensation control submodule is configured to generate a light compensation device angle deflection instruction according to the coordinate distribution of the low contrast region, dynamically adjust the light compensation intensity level in combination with water turbidity data, and output an optimized visible light image sequence. 9.The image fusion-based underwater fish school monitoring and counting system according to claim 1, wherein, The system further includes: The data correction module is configured to receive the fish school number distribution statistical thermal map of the population statistical output module, extract population density benchmark values in historical same-period monitoring data, calculate a deviation rate of current distribution data from the benchmark values, perform spatial coordinate correction on regions with a deviation rate exceeding a threshold, and output a corrected fish school statistical distribution map.

Images