A device and method for fish behavior monitoring

Abstract

The present application relates to the technical field of image processing, in particular to a device and method for fish behavior monitoring, the device comprising a fish body grid construction module, a vertex optical flow tracking module, a compressed potential energy calculation module, an energy field thermal mapping module and a swimming mode identification module. In the present application, a triangular topological grid covering the surface of the organism is constructed and the time sequence vertex optical flow trajectory is tracked, the area scaling ratio of the local grid unit in the movement process is analyzed by using the affine transformation matrix, the geometric deformation in the visual level is mapped into the physical potential energy index representing the muscle contraction intensity, the small deformation difference is amplified by combining the nonlinear index operation, the visual energy distribution thermal map is generated by Gaussian smoothing processing, the conduction characteristics of the high-energy connected region of the body in the space-time dimension are identified, the leap from simple trajectory observation to biomechanical function evaluation is realized, and the energy consumption efficiency and burst starting characteristics under different tail wagging modes can be accurately identified.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and in particular to a device and method for monitoring fish behavior. Background Technology

[0002] Image processing technology primarily involves the process of processing and analyzing image information using computers and dedicated hardware systems. Its core aspects include image digitization, encoding, compression, enhancement, restoration, segmentation, and feature extraction. By performing linear or nonlinear mathematical operations on the image pixel matrix, analog light signals are converted into digital signals, and arithmetic logic units are used to transform the image data to analyze the visual features of the object. Traditionally, devices used for monitoring fish behavior refer to hardware systems based on machine vision principles that perform non-contact observation of underwater fish movement. These systems employ underwater cameras or industrial cameras as acquisition terminals, transmitting video streams to a host computer via coaxial cables or network cables. A video capture card discretizes the analog signals into digital image frames, and a background model is constructed using background subtraction or inter-frame subtraction methods. The foreground motion region is extracted by calculating the pixel grayscale difference between the current frame and the background model, and the position and trajectory of the fish in the image coordinate system are obtained by calculating the centroid of the connected components.

[0003] Existing technologies mostly rely on underwater camera equipment to acquire images and use background subtraction to extract the foreground region. The motion trajectory is obtained by calculating the centroid of the connected domain. This observation mode based on geometric mass points can only depict the macroscopic displacement path and cannot capture the subtle deformations caused by muscle contraction inside the fish. It is difficult to quantify the instantaneous explosive force and work intensity at the biomechanical level. As a result, the monitoring of fish tail wagging patterns and physiological functions remains at the level of shallow behavioral trajectory. It lacks deep quantitative evidence that reflects the relaxation state of the body muscles and energy transfer efficiency, and cannot meet the needs of refined behavioral analysis. Summary of the Invention

[0004] To address the limitations of existing technologies, which can only depict macroscopic displacement paths and cannot capture the subtle deformations caused by muscle contraction within the fish body, making it difficult to quantify the instantaneous explosive force and work intensity at the biomechanical level, resulting in monitoring of fish tail-wagging patterns and physiological functions remaining at the superficial behavioral trajectory level and lacking deep quantitative evidence reflecting the relaxation state of the body's muscles and energy transfer efficiency, thus failing to meet the technical problems of refined behavioral analysis, this invention provides a device and method for monitoring fish behavior. The technical solution is as follows: On the one hand, a device for monitoring fish behavior is provided, the device comprising: The fish body mesh construction module collects continuous underwater video stream data of the target fish, scans pixel grayscale gradient changes, identifies the contour range, extracts the geometric midline to define the spinal axis, filters feature positioning points and connects them, and generates a set of triangular mesh vertices on the fish body surface. The vertex optical flow tracing module calls the vertex set of the triangular mesh on the surface of the fish, retrieves the same texture feature pixel region in adjacent time frames, matches the evolution trajectory of vertex spatial position, and establishes temporal evolution vertex coordinate mapping pairs. The compression potential energy calculation module constructs an affine transformation matrix from the previous frame to the real frame based on the temporal evolution vertex coordinate mapping pair, calculates the local area scaling Jacobian determinant, and performs differential operations and nonlinear exponential amplification in combination with rigid body motion reference parameters to obtain the local mesh compression potential energy index. The energy field thermodynamic mapping module maps the local grid compression potential energy index to the fish body topology, performs weighted smooth interpolation, scans the continuous energy distribution spectrum, identifies energy intensity clusters exceeding preset limits, and generates high-energy connected domains for body work. The swimming pattern identification module calculates the position of the high-energy connected domain of the body's work at the geometric center of the spinal axis, analyzes the conduction characteristics and spatial distribution pattern, matches the biomechanical motion model to determine the type, and obtains the classification results of fish tail-wagging swimming posture.

[0005] As a further aspect of the present invention, the triangular mesh vertex set on the fish body surface includes node three-dimensional coordinate data, a triangular patch index table, and vertex normal vectors; the temporal evolution vertex coordinate mapping pair includes inter-frame displacement correlation vectors, timestamp sequence markers, and trajectory correlation coefficients; the local mesh compression potential energy index includes area scaling deformation rate, nonlinear potential energy value, and muscle contraction state identifier; the high-energy connected domain of body work includes the outline of energy accumulation region, high-energy core distribution coordinates, and extreme values ​​of regional work intensity; and the fish tail-wagging swimming posture classification result includes swimming posture pattern category label, movement mechanism matching degree, and energy consumption efficiency evaluation parameters.

[0006] As a further aspect of the present invention, the fish body mesh construction module includes: The contour scanning and recognition submodule acquires continuous underwater video stream data of the target fish, performs pixel-level scanning operation on a single frame image matrix, calculates the gray-level gradient magnitude and direction vector between adjacent pixels, defines the boundary line between the foreground target and the background area based on the gradient change position, extracts continuous edge pixel coordinates and arranges them in a clockwise direction, identifies the geometric shape range of the organism in the image coordinate system, and generates a fish body edge contour coordinate sequence. The feature node extraction submodule calls the fish body edge contour coordinate sequence, traverses the pixels inside the contour and the boundary region, calculates the corner response function value and edge curvature extreme value of the local region, retains the pixel points whose response values ​​exceed the feature screening benchmark as key nodes, performs non-maximum suppression operation, removes collinear points and discrete noise points, and obtains the set of feature positioning points on the surface of the organism. The topology mesh generation submodule performs Delaunay triangulation based on the set of feature points on the surface of the organism, connects the three feature points with the closest spatial distance according to the empty circle property criterion, constructs non-overlapping triangular units, maximizes the minimum angle of the interior angle of the triangle, calculates the vertex index relationship and adjacency list information of the triangular units, and obtains the vertex set of the triangular mesh on the surface of the fish.

[0007] As a further aspect of the present invention, the vertex optical flow tracing module includes: The texture feature retrieval submodule calls the triangular mesh vertex set on the fish body surface to obtain video frame image data at the real time and the next adjacent time. It constructs a local pixel window with the mesh vertex coordinates as the center, calculates the gray-level distribution statistical features and texture gradient direction of the pixels in the window, performs cross-correlation operation or least squares difference matching within the search range of adjacent time frame images, identifies biological tissue pixel regions with the same texture gray-level features, and generates an index of texture matching regions in adjacent frames. The spatial trajectory matching submodule calculates the pixel coordinate displacement of the grid vertex in the continuous time series based on the adjacent frame texture matching region index, performs optical flow constraint operation, solves the instantaneous motion velocity vector of the vertex on the image plane, corrects the motion direction of abnormal drift points by combining smoothing constraint conditions, connects the spatial position coordinates of the same vertex at different times, matches the spatial position evolution trajectory of the vertex at different times, and obtains the vertex spatiotemporal displacement vector sequence. The temporal mapping construction submodule extracts the two-dimensional coordinate values ​​and corresponding timestamps of the vertices based on the spatiotemporal displacement vector sequence of the vertices, arranges the spatial position data of the vertices in chronological order, constructs an association index between the unique identifier of the vertex and the spatiotemporal coordinate data, identifies the motion migration relationship of discrete nodes on the surface of the organism in the time dimension, and establishes a temporal evolution vertex coordinate mapping pair.

[0008] As a further aspect of the present invention, the compression potential energy calculation module includes: The affine transformation construction submodule extracts the vertex coordinate groups of the same triangular unit in the previous frame and the real frame based on the temporal evolution vertex coordinate mapping pair, establishes a linear transformation equation describing the evolution of the geometric shape from the previous moment to the real moment, analyzes the equation coefficients using the least squares method, determines the affine parameters including rotation, shearing and translation components, constructs a mathematical transformation model that maps the triangular geometric shape of the previous frame to the shape of the real frame, and generates a local geometric shape transformation matrix. The area scaling calculation submodule calls the local geometric transformation matrix, extracts the diagonal elements representing the scaling ratio of the coordinate axis direction and the off-diagonal elements representing the degree of inclination between axes, performs second-order determinant algebra operation, analyzes the scaling ratio of the local mesh unit area micro-element, calculates the Jacobian determinant value representing the local area scaling state, quantifies the degree of geometric deformation of the triangular unit, and obtains the local area scaling determinant value. The potential energy calculation submodule, based on the local area scaling determinant value, introduces the rigid body motion reference parameter characterizing the non-deformation state, performs numerical difference operation, calculates the deviation between the real-time grid area change rate and the reference value, performs nonlinear exponential amplification processing on the difference result, optimizes the numerical resolution of the deformation range, and obtains the local grid compression potential energy index.

[0009] As a further aspect of the present invention, the energy field thermodynamic mapping module includes: The energy topology mapping submodule calls the local mesh compression potential energy index, maps the quantized values ​​one by one to the spatial geometric nodes defined by the triangular mesh vertex set on the fish body surface, constructs a discrete scalar field characterizing the force distribution on the body surface, calculates the spatial distance weights of nodes in the local neighborhood relative to the center point, performs weighted smooth interpolation processing, fills the mesh gaps and suppresses noise fluctuations, and establishes a continuous energy distribution map. The intensity limit screening submodule scans the global grid node values ​​included in the continuous energy distribution map, introduces a preset energy intensity limit benchmark parameter, filters active nodes whose node values ​​exceed the energy intensity limit benchmark parameter, records spatial coordinates and filters background noise areas to obtain the set of pixels with excess energy. The connected component morphology extraction submodule calls the set of super-energy pixels, analyzes the neighborhood adjacency state of discrete pixels in the image space, performs morphological closing operation, bridges gaps and merges adjacent isolated point clusters, and aggregates spatially continuous pixels into independent regions according to connectivity rules to generate high-energy connected components for body work.

[0010] As a further aspect of the present invention, the process of introducing a preset energy intensity limit reference parameter specifically involves: statistically analyzing the energy scalar values ​​of discrete nodes in a continuous energy distribution spectrum, calculating the arithmetic mean of the energy values ​​across the entire domain as a reference background level, calculating the root mean square deviation of the energy scalar values ​​of discrete nodes relative to the arithmetic mean, and quantifying the degree of dispersion of the energy distribution across the entire domain. Set a confidence factor greater than 0, multiply the confidence factor by the root mean square deviation, and obtain the noise fluctuation tolerance. The arithmetic mean and the noise fluctuation tolerance are added together to output a uniquely determined preset energy intensity limit reference parameter.

[0011] As a further aspect of the present invention, the swimming pattern recognition module includes: The spinal center localization submodule calls the high-energy connected domain of the body to extract the planar coordinate distribution data of the pixels inside the connected region, calculates the first-order geometric moment of the region, determines the position of the centroid on the image plane, introduces the linear equation parameters of the fish spinal axis, performs the projection operation from point to line, maps the centroid coordinates to the direction of the spinal axis, quantifies the distribution position of the high-energy region on the body trunk, and obtains the spinal projection coordinates of the high-energy region. The spatiotemporal feature analysis submodule, based on the spinal projection coordinates of the high-energy region, arranges the projection position data in consecutive frames in a time sequence, calculates the instantaneous velocity vector and acceleration change rate of the coordinate point moving along the spinal direction, analyzes the phase delay relationship transmitted from the head to the tail, calculates the eccentricity and compactness parameters in combination with the spatial distribution range of the connected domain, analyzes the spatiotemporal evolution law of the work area of ​​the organism, and generates the spatiotemporal conduction feature vector of energy flow. The motion model matching submodule, based on the energy flow spatiotemporal conduction feature vector, calls the feature parameter set of standard swimming strokes in the preset biomechanical motion model library, performs similarity calculation or Euclidean distance measurement operation in the vector space, finds the benchmark model category that matches the real-time feature vector, determines the motion mechanism type of the target object in the real-time observation period, and obtains the classification result of fish tail-wagging swimming stroke.

[0012] As a further aspect of the present invention, the process of calculating the instantaneous velocity vector and acceleration rate of change of the coordinate point moving along the spine direction is specifically as follows: extracting the high-energy region spine projection coordinates corresponding to the real-time sampling frame and the previous sampling frame, calculating the algebraic difference between the coordinate values ​​of the two time points in the direction of the spine axis, and obtaining the displacement increment. Obtain the frame rate parameters of the video acquisition device and calculate the inter-frame time interval. Divide the displacement increment by the inter-frame time interval to obtain the instantaneous velocity vector characterizing the first derivative. Extract instantaneous velocity vector data from three consecutive time points, calculate the change in the difference between adjacent velocity vectors, divide the change in the difference by the inter-frame time interval, and obtain the acceleration value representing the second derivative.

[0013] On the other hand, a method for monitoring fish behavior, wherein the method is performed based on the above-mentioned fish behavior monitoring device, includes the following steps: S1: Collect continuous underwater video stream data of the target fish, scan the pixel grayscale gradient changes within the video frame, identify the fish body outline range, filter the feature positioning points within the outline and connect them to generate a set of triangular mesh vertices on the fish body surface. S2: Call the triangular mesh vertex set on the surface of the fish, retrieve pixel regions with the same texture features in adjacent time frames, match the evolution trajectory of vertices in spatial position, and establish temporal evolution vertex coordinate mapping pairs; S3: Based on the temporal evolution vertex coordinate mapping pair, construct the morphological affine transformation matrix from the previous frame to the real frame, use the matrix elements to calculate the local area scaling Jacobian determinant, combine the rigid body motion reference parameters to perform differential operation on the Jacobian determinant, and perform nonlinear exponential amplification processing on the differential operation value to obtain the local mesh compression potential energy index. S4: Map the local mesh compression potential energy index to the fish body topology, perform weighted smooth interpolation, scan the continuous energy distribution spectrum, identify the clustered regions where the energy intensity exceeds the preset limit, and generate the high-energy connected domain for body work. S5: Calculate the position of the high-energy connected domain of the body doing work at the geometric center of the spinal axis, analyze the conduction characteristics and spatial distribution of the connected domain, match the biomechanical motion model parameters to determine the type, and obtain the classification results of fish tail-wagging swimming posture.

[0014] The beneficial effects of the technical solutions provided in the embodiments of the present invention include at least the following: By constructing a triangular topological mesh covering the surface of a living organism and tracing the temporal vertex optical flow trajectory, the area scaling ratio of local mesh cells during motion is analyzed using an affine transformation matrix. This maps the geometric deformation at the visual level into a physical potential energy index characterizing the intensity of muscle contraction. By combining nonlinear exponential operations to amplify the differences in minute deformations, and generating a visualized energy distribution thermogram through Gaussian smoothing, the transmission characteristics of high-energy connected regions in the body in the spatiotemporal dimension are identified. This achieves a leap from simple trajectory observation to biomechanical function assessment, and can accurately identify the energy consumption efficiency and burst initiation characteristics in different tail-wagging modes. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 This is a schematic diagram of the device provided by the present invention; Figure 2 This is a schematic diagram of the device frame of the present invention; Figure 3 This is a flowchart of the fish body mesh construction module in this invention; Figure 4 This is a flowchart of the vertex optical flow tracing module in this invention; Figure 5 This is a flowchart of the compression potential energy calculation module in this invention; Figure 6 This is a flowchart of the energy field thermodynamic mapping module in this invention; Figure 7 This is a flowchart of the movement pattern recognition module in this invention; Figure 8 This is a flowchart of the method provided by the present invention. Detailed Implementation

[0017] The technical solution of the present invention will now be described with reference to the accompanying drawings.

[0018] To make the technical problems, technical solutions and advantages of the present invention clearer, a detailed description will be given below in conjunction with the accompanying drawings and specific embodiments.

[0019] This invention provides a device for monitoring fish behavior, such as... Figure 1-2 The schematic diagram shown is for a fish behavior monitoring device, which includes: The fish body mesh construction module monitors fish movement through miniature sensors worn on the fish body, collects continuous underwater video stream data of the target fish, scans the gray-scale gradient changes of image pixels, identifies the range of the organism's edge contour, extracts the geometric central axis to define the spinal axis, filters feature positioning points inside the contour and boundary areas, connects the feature positioning points, and generates a set of triangular mesh vertices on the fish body surface. The vertex optical flow tracing module calls the vertex set of the triangular mesh on the fish surface, retrieves the biological tissue pixel regions with the same texture grayscale features in adjacent time frame images, matches the evolution trajectory of the spatial position of vertices at different times, identifies the movement and migration relationship of discrete nodes on the surface of the organism in the time dimension, and establishes temporal evolution vertex coordinate mapping pairs. The compression potential energy calculation module constructs an affine transformation matrix that maps the triangular geometry of the previous frame to the shape of the real frame based on the vertex coordinate mapping pairs of the temporal evolution. It calculates the Jacobian determinant that represents the local area scaling state, and performs differential operations and nonlinear exponential amplification processing in combination with the rigid body motion reference parameters to quantify the muscle contraction deformation intensity and obtain the local mesh compression potential energy index. The energy field thermodynamic mapping module maps the local mesh compression potential energy index to the fish body topology, performs weighted smoothing interpolation operation between adjacent nodes to avoid discrete noise, scans the smoothed continuous energy distribution spectrum, identifies energy intensity clusters exceeding the preset limit and extracts morphological features, and generates high-energy connected domains for body work. The swimming pattern identification module calculates the geometric center position of the high-energy connected domain of the body's work along the direction of the fish's spinal axis, analyzes the temporal transmission characteristics and spatial distribution of the high-energy connected domain, matches it with a preset biomechanical motion model to determine the type, and obtains the classification results of the fish's tail-wagging swimming posture. The fish body surface triangular mesh vertex set includes node 3D coordinate data, triangular patch index table, and vertex normal vector. The temporal evolution vertex coordinate mapping pair includes inter-frame displacement correlation vector, timestamp sequence marker, and trajectory correlation coefficient. The local mesh compression potential energy index includes area scaling deformation rate, nonlinear potential energy value, and muscle contraction state identifier. The high-energy connected domain of body work includes the outline of energy accumulation region, high-energy core distribution coordinates, and extreme values ​​of regional work intensity. The fish tail-wagging swimming posture classification results include swimming posture pattern category label, movement mechanism matching degree, and energy consumption efficiency evaluation parameters.

[0020] Specifically, such as Figure 2 , 3 As shown, the fish body mesh building module includes: The contour scanning and recognition submodule acquires continuous underwater video stream data of the target fish, performs pixel-level scanning operation on a single frame image matrix, calculates the gray-level gradient magnitude and direction vector between adjacent pixels, defines the boundary line between the foreground target and the background area based on the gradient change position, extracts continuous edge pixel coordinates and arranges them in a clockwise direction, identifies the geometric shape range of the organism in the image coordinate system, and generates a fish body edge contour coordinate sequence. Equipped with a high-speed underwater vision acquisition unit, it acquires continuous underwater video stream data at a resolution of 3840 x 2160 pixels at a rate of 60 frames per second, and converts it into a single-frame RGB image matrix via a built-in decoder. The gradient calculation unit within the submodule performs pixel-level scanning operations on the single-frame image matrix, calling a preset Sobel operator convolution kernel to calculate the gray-level difference values ​​of the target pixel along the horizontal and vertical coordinate axes. The gradient calculation unit further performs arithmetic operations, adding the squares of the horizontal and vertical gradients and taking the square root to obtain the gray-level gradient magnitude of the pixel. The direction vector is determined by calculating the arctangent function of the ratio of the vertical to the horizontal gradient. The submodule's boundary determination unit has a built-in gradient threshold register with a value of 128. When the calculated gradient magnitude is greater than this threshold, the pixel is determined to be the boundary line between the foreground target and the background region. Subsequently, the edge tracking unit extracts the coordinates of all consecutive edge pixels that meet the conditions and arranges them in a clockwise direction to identify the geometric shape range of the organism in the image coordinate system, and finally outputs the fish body edge contour coordinate sequence.

[0021] The feature node extraction submodule calls the fish body edge contour coordinate sequence, traverses the pixels inside the contour and the boundary region, calculates the corner response function value and edge curvature extreme value of the local region, retains the pixels whose response value exceeds the feature screening benchmark as key nodes, performs non-maximum suppression operation, removes collinear points and discrete noise points, and obtains the set of feature localization points on the surface of the organism. This submodule directly connects to the data interface of the contour scanning and recognition submodule, calling the fish body edge contour coordinate sequence. The response calculation unit within the submodule traverses the pixels within the contour and boundary regions, constructing a 5x5 pixel local detection window and calculating the corner response function values ​​of the local region. This calculation involves subtracting the product of the square of the trace of the structure tensor matrix and the empirical constant 0.04 from the determinant value of the structure tensor matrix. Simultaneously, the curvature analysis unit calculates the gradient rate of change along the edge tangent direction to obtain the edge curvature extrema. The filtering unit compares the calculated response values ​​with a preset feature filtering benchmark of 1000, retaining only pixels with response values ​​exceeding this benchmark. To optimize the point set distribution, the non-maximum suppression unit performs suppression operations within a 10-pixel radius neighborhood centered on the retained point. If a point with a higher response value is detected, the current point is discarded, effectively eliminating collinear points and discrete noise points to obtain a highly reliable set of biological surface feature localization points.

[0022] The topology mesh generation submodule is based on the set of feature localization points on the surface of the organism. It performs Delaunay triangulation operation, connects the three feature points with the closest spatial distance according to the empty circle property criterion, constructs non-overlapping triangular units, maximizes the minimum angle of the interior angle of the triangle, calculates the vertex index relationship and adjacency list information of the triangular units, and obtains the vertex set of the triangular mesh on the surface of the fish. The spatial triangulation task is performed. The Delaunay triangulation engine built into the submodule follows the empty circle property criterion, ensuring that the circumcircle of any generated triangle does not contain any other feature points in the point set. The connection unit calculates the Euclidean distance between the three closest feature points in space using an iterative algorithm, and connects the three points that satisfy the empty circle condition to construct non-overlapping triangular units. To ensure mesh quality, the optimization unit performs an edge swap operation, maximizing the minimum angle of the triangle's interior angle to ensure that the minimum interior angle of all generated units is greater than 30 degrees, thereby avoiding the impact of the elongated shape on the stability of numerical computation. Finally, the mesh index unit calculates and stores the index relationship of the three vertices of each triangular unit and the corresponding adjacency list information, outputting the complete vertex set of the triangular mesh on the fish's surface.

[0023] Specifically, such as Figure 2 , 4 As shown, the vertex optical flow tracing module includes: The texture feature retrieval submodule calls the triangular mesh vertex set on the fish body surface to obtain video frame image data at the real time and the next adjacent time. It constructs a local pixel window with the mesh vertex coordinates as the center, calculates the gray-level distribution statistical features and texture gradient direction of the pixels within the window, performs cross-correlation operation or least squares difference matching within the search range of adjacent time frame images, identifies biological tissue pixel regions with the same texture gray-level features, and generates an index of texture matching regions in adjacent frames. The submodule connects to the video frame buffer to obtain video frame image data from the real-time moment and the next adjacent moment. The window construction unit within the submodule constructs a local pixel window with a side length of 15 pixels, centered on the grid vertex coordinates. The feature statistics unit calculates the distribution statistics of the pixels within the window, including the mean and variance of grayscale values, as well as the texture gradient direction. The matching operation unit performs normalized cross-correlation within a preset search range of adjacent time frame images, such as a region with a radius of 30 pixels. This operation calculates the similarity coefficient by summing the products of the corresponding pixel grayscale values ​​in the template window and the search window, and dividing by the square root of the product of their autocorrelation coefficients. Based on the operation results, the submodule identifies the region with the highest correlation coefficient as a biological tissue pixel region with the same texture grayscale features, and the index generation unit generates an index of adjacent frame texture matching regions containing positional correspondences.

[0024] The spatial trajectory matching submodule calculates the pixel coordinate displacement of the grid vertex in the continuous time series based on the texture matching region index of adjacent frames, performs optical flow constraint operation, solves the instantaneous motion velocity vector of the vertex on the image plane, corrects the motion direction of abnormal drift points by combining smoothing constraint conditions, connects the spatial position coordinates of the same vertex at different times, matches the spatial position evolution trajectory of the vertex at different times, and obtains the vertex spatiotemporal displacement vector sequence. The displacement calculation unit calculates the pixel coordinate displacement of mesh vertices in a continuous time series. The optical flow calculation unit performs optical flow constraint calculations, establishing a system of linear equations containing unknown horizontal and vertical velocities based on the assumption of constant brightness, and solving it using the least squares method to calculate the instantaneous velocity vector of the vertex on the image plane. To eliminate tracking errors, the smoothing correction unit, combined with smoothing constraints, calculates the weighted average of the velocity vectors of other vertices in the neighborhood of the target vertex. When the deviation of the target vertex's velocity vector from this average exceeds a set threshold of 2 pixels per frame, the average is used to correct the motion direction of the abnormal drift point. The trajectory connection unit is responsible for connecting the spatial coordinates of the same vertex at differentiated times, matching the spatial position evolution trajectory of the vertex at differentiated times, and outputting the spatiotemporal displacement vector sequence of the vertex.

[0025] The temporal mapping construction submodule extracts the two-dimensional coordinate values ​​and corresponding timestamps of vertices based on the spatiotemporal displacement vector sequence of vertices, arranges the spatial position data of vertices in chronological order, constructs an association index between the unique identifier of the vertex and the spatiotemporal coordinate data, identifies the motion migration relationship of discrete nodes on the surface of organisms in the time dimension, and establishes temporal evolution vertex coordinate mapping pairs. The coordinate extraction unit extracts the two-dimensional coordinate values ​​of vertices and their corresponding timestamps. The data arrangement unit queues the spatial location data of vertices according to their chronological order. The index construction unit is responsible for building an index that associates unique vertex identifiers with spatiotemporal coordinate data, ensuring the uniqueness of each physical node's identity on the timeline. By identifying the movement and migration relationships of discrete nodes on the surface of a biological organism in the time dimension, this submodule establishes a temporal evolution vertex coordinate mapping pair containing vertex ID, timestamp t, x-coordinate, and y-coordinate.

[0026] Specifically, such as Figure 2 , 5 As shown, the compression potential energy calculation module includes: The affine transformation construction submodule extracts the vertex coordinate sets of the same triangular unit in the previous frame and the real frame based on the vertex coordinate mapping pairs of the temporal evolution, establishes a linear transformation equation describing the evolution of the geometric shape from the previous time to the real time, analyzes the coefficients of the equation using the least squares method, determines the affine parameters including rotation, shearing and translation components, constructs a mathematical transformation model that maps the triangular geometric shape of the previous frame to the shape of the real frame, and generates a local geometric shape transformation matrix. This submodule, based on the temporal evolution vertex coordinate mapping pairs, extracts the vertex coordinate sets of the same triangular unit in the previous frame and the real-time frame using the coordinate extraction unit. The model construction unit establishes a linear transformation equation describing the evolution of the geometric shape from the previous time step to the real-time time step. The parameter analysis unit uses the least squares method to analyze the coefficients of the equation and determines the affine parameters, including the cosine value of the rotation angle, the shearing scale factor, and the translation distance component, through matrix operations. This process constructs a mathematical transformation model that maps the triangular geometry of the previous frame to the shape of the real-time frame, and the matrix generation unit generates a local geometric transformation matrix containing six parameters. For example, for the reference frame coordinate set and the real-time frame coordinate set, the submodule calculates that the main diagonal elements of the transformation matrix are 1.05 and 0.98, respectively, representing the stretching and compression states of the local region in two axes.

[0027] The area scaling calculation submodule calls the local geometric transformation matrix, extracts the diagonal elements representing the scaling ratio of the coordinate axis direction and the off-diagonal elements representing the degree of inclination between axes, performs second-order determinant algebra operation, analyzes the scaling ratio of the local mesh cell area micro-element, calculates the Jacobian determinant value representing the local area scaling state, quantifies the degree of geometric deformation of the triangular cell, and obtains the local area scaling determinant value. The determinant operation unit extracts the diagonal elements representing the scaling ratio along the coordinate axes and the off-diagonal elements representing the tilt between axes within the matrix. The operation unit performs second-order determinant algebra, calculating the product of the two elements on the main diagonal and subtracting the product of the two elements on the secondary diagonal to analyze the scaling ratio of the local mesh cell area. This result is the Jacobian determinant value representing the local area scaling state. The quantization unit outputs this value to quantify the degree of geometric deformation of the triangular cells, obtaining the local area scaling determinant value. If this value is less than 1, it indicates that compression deformation has occurred in the local region; if it is greater than 1, it indicates that expansion deformation has occurred.

[0028] The potential energy calculation submodule is based on the local area scaling determinant value. It introduces the rigid body motion reference parameter to characterize the non-deformation state, performs numerical difference operation, calculates the deviation between the real-time grid area change rate and the reference value, performs nonlinear exponential amplification processing on the difference result, optimizes the numerical resolution of the deformation range, and obtains the local grid compression potential energy index. It has a built-in rigid body motion reference parameter register with a value of 1, representing the non-deformation state. The difference operation unit performs numerical difference operations to calculate the deviation between the real-time mesh area change rate and the reference value, that is, to calculate the absolute value of the difference between the determinant value and the reference value. The exponential processing unit performs nonlinear exponential amplification processing on the difference results, by subtracting 1 from the power of the difference result of the natural constant e, or by multiplying the square of the difference result by the preset elastic modulus coefficient 100, to optimize the numerical resolution of the deformation range. Finally, this submodule outputs a local mesh compression potential energy index, which can significantly amplify small deformation differences, thereby accurately locating high-stress areas; [Logical Deduction and Parameter Calculation] To verify the logical validity of the compression potential energy calculation module, the following derivation is performed through a practical example: Parameter acquisition: The area scaling calculation submodule outputs a Jacobian determinant value of 0.85 for a local element, indicating that the region is under compression. The preset rigid body motion reference parameter is 1. The elastic modulus is set to 200. Logical operation: The differential operation unit calculates the degree of deviation: subtract 1 from 0.85 and take the absolute value to get 0.15.

[0029] Perform square amplification and weighting: Calculate the square of 0.15 to get 0.0225; then multiply 0.0225 by the elastic modulus coefficient 200 to get 4.5.

[0030] Results Analysis: The final calculated local mesh compression potential energy index is 4.5. If the determinant value of another region is 0.98, the potential energy index is calculated using the same logic (0.02 squared multiplied by 200) to be 0.08. By comparison, 4.5 is much greater than 0.08, indicating that the numerical results show that the first region is subjected to a significant compression load, and the device will determine it as the main work region based on this.

[0031] Specifically, such as Figure 2 , 6 As shown, the energy field thermodynamic mapping module includes: The energy topology mapping submodule calls the local mesh compression potential energy index, maps the quantized values ​​one by one to the spatial geometric nodes defined by the triangular mesh vertex set on the fish surface, constructs a discrete scalar field characterizing the force distribution on the body surface, calculates the spatial distance weights of nodes in the local neighborhood relative to the center point, performs weighted smooth interpolation processing, fills the mesh gaps and suppresses noise fluctuations, and establishes a continuous energy distribution map. The mapping unit maps quantized values ​​one-to-one to spatial geometric nodes defined by the vertex set of the triangular mesh on the fish's surface, constructing a discrete scalar field representing the force distribution on the body surface. The interpolation unit is responsible for calculating the spatial distance weights of nodes in the local neighborhood relative to the center point. The weight value is set as the reciprocal of the square of the Euclidean distance from that point to the center point. The smoothing unit performs weighted smoothing interpolation. For any pixel in the mesh gap, it selects the 5 nearest feature nodes, calculates the sum of the products of the potential energy index of these 5 nodes and their corresponding weights, and divides by the sum of all weights. This fills the mesh gaps and suppresses noise fluctuations, establishing a continuous and smooth continuous energy distribution map to ensure the spatial continuity of the energy field.

[0032] The intensity limit screening submodule scans the global grid node values ​​included in the continuous energy distribution map, introduces the preset energy intensity limit benchmark parameters, filters active nodes whose node values ​​exceed the energy intensity limit benchmark parameters, records spatial coordinates and filters background noise areas to obtain the set of pixels with excess energy. This submodule is equipped with a global scanning unit for scanning the values ​​of all grid nodes in the continuous energy distribution map. The parameter setting unit introduces a preset energy intensity limit benchmark parameter, which is dynamically determined by statistically analyzing the top 15% quantiles of the global energy distribution histogram. The filtering unit compares the energy value of each node with the energy intensity limit benchmark parameter, filtering out active nodes whose values ​​exceed the benchmark. The coordinate recording unit records the spatial coordinates of these active nodes and filters out background noise areas with values ​​below the benchmark, thereby obtaining a set of pixels exceeding the energy limit and achieving preliminary segmentation of the high-energy work region.

[0033] The connected component morphology extraction submodule calls the set of super-energy pixels, analyzes the neighborhood adjacency state of discrete pixels in the image space, performs morphological closing operation, bridges gaps and merges adjacent isolated point clusters, and aggregates spatially continuous pixels into independent regions according to connectivity rules to generate high-energy connected components for body work. The neighborhood analysis unit analyzes the adjacency state of discrete pixels in the image space. The morphological processing unit performs morphological closing operations. First, it performs a dilation operation using a 3x3 pixel structuring element to expand the boundary of high-energy regions to fill small gaps. Then, it performs an erosion operation to restore the original size of the regions, thereby bridging gaps and merging adjacent isolated point clusters. The clustering unit aggregates spatially continuous pixels into independent regions according to the 8-neighborhood connectivity rule, generating a defined high-energy connected domain for body work, thus clarifying the physical location where the organism actually generates thrust.

[0034] Specifically, such as Figure 2 , 7 As shown, the swimming pattern recognition module includes: The spinal center localization submodule calls the high-energy connected domain of the body's work, extracts the planar coordinate distribution data of the pixels inside the connected region, calculates the first-order geometric moment of the region, determines the position of the centroid on the image plane, introduces the linear equation parameters of the fish's spinal axis, performs the projection operation from point to line, maps the centroid coordinates to the direction of the spinal axis, quantifies the distribution position of the high-energy region on the organism's trunk, and obtains the spinal projection coordinates of the high-energy region. The planar coordinate distribution data of pixels within the connected region is extracted. The centroid calculation unit calculates the first-order geometric moment of the region, that is, the sum of the abscissas of all pixels in the region is divided by the total number of pixels to obtain the abscissa of the centroid. The ordinate is calculated similarly to determine the position of the centroid on the image plane. The projection operation unit introduces the linear equation parameters of the fish's spinal axis and performs a point-to-line projection operation to calculate the perpendicular distance from the centroid coordinates to the spinal axis and the coordinates of the foot of the perpendicular, mapping the centroid coordinates to the direction of the spinal axis. The quantization unit calculates the distance between the foot of the perpendicular and the feature point of the fish head and divides it by the total length of the spine to quantize the distribution position of the high-energy region on the organism's torso, obtaining the normalized spinal projection coordinates of the high-energy region.

[0035] The spatiotemporal feature analysis submodule is based on the high-energy region spinal projection coordinates. It arranges the projection position data in consecutive frames in a time sequence, calculates the instantaneous velocity vector and acceleration change rate of the coordinate point moving along the spinal direction, analyzes the phase delay relationship transmitted from the head to the tail, calculates the eccentricity and compactness parameters in combination with the spatial distribution range of the connected domain, analyzes the spatiotemporal evolution law of the work area of ​​the organism, and generates the spatiotemporal conduction feature vector of energy flow. The projection position data in consecutive frames are arranged in a time sequence. The motion analysis unit calculates the instantaneous velocity vector of the coordinate point moving along the spine direction, i.e., the difference between the projection positions of adjacent frames divided by the frame interval, and simultaneously calculates the rate of change of acceleration. The phase analysis unit analyzes the phase delay relationship transmitted from the head to the tail, and calculates the peak delay of the cross-correlation between the head swing signal and the high-energy region movement signal. The morphology analysis unit calculates the eccentricity and compactness parameters (the ratio of area to the square of the perimeter) based on the spatial distribution range of the connected domains, comprehensively analyzes the spatiotemporal evolution law of the organism's work area, and generates a spatiotemporal energy flow conduction feature vector containing velocity, phase, and morphological parameters.

[0036] The motion model matching submodule uses the energy flow spatiotemporal conduction feature vector to call the feature parameter set of standard swimming strokes in the preset biomechanical motion model library, performs similarity calculation or Euclidean distance measurement operation in the vector space, finds the benchmark model category that matches the real-time feature vector, determines the motion mechanism type of the target object in the real-time observation period, and obtains the classification result of fish tail wagging swimming stroke. The system calls upon a pre-defined biomechanical motion model library containing feature parameters of standard swimming strokes. This library includes standard pattern vectors for fish such as trevally and eels. The similarity calculation unit performs similarity calculations in the vector space, using a cosine similarity algorithm to calculate the cosine of the angle between the real-time feature vector and each standard template vector. The classification unit searches for a baseline model category that matches the real-time feature vector. If the calculated similarity value exceeds a pre-defined threshold of 0.85, the system determines the motion mechanism type of the target object during the real-time observation period, obtaining the fish tail-wagging swimming stroke classification result.

[0037] Please see Figure 8 The method for monitoring fish behavior is based on the above-mentioned fish behavior monitoring device and includes the following steps: S1: Collect continuous underwater video stream data of the target fish, scan the pixel grayscale gradient changes within the video frame, identify the fish body outline range, filter the feature positioning points within the outline and connect them to generate a set of triangular mesh vertices on the fish body surface. S2: Call the triangular mesh vertex set on the surface of the fish, retrieve pixel regions with the same texture features in adjacent time frames, match the evolution trajectory of vertices in spatial position, and establish temporal evolution vertex coordinate mapping pairs; S3: Based on the vertex coordinate mapping pairs of the temporal evolution, construct the morphological affine transformation matrix from the previous frame to the real frame, use the matrix elements to calculate the local area scaling Jacobian determinant, combine the rigid body motion reference parameters to perform difference operations on the Jacobian determinant, and perform nonlinear exponential amplification processing on the difference operation value to obtain the local mesh compression potential energy index. S4: Map the local mesh compression potential energy index to the fish body topology, perform weighted smooth interpolation, scan the continuous energy distribution spectrum, identify the clustered regions where the energy intensity exceeds the preset limit, and generate the high-energy connected domains where the body does work. S5: Calculate the location of the high-energy connected domain of the body doing work at the geometric center of the spinal axis, analyze the conduction characteristics and spatial distribution of the connected domain, match the parameters of the biomechanical motion model to determine the type, and obtain the classification results of fish tail-wagging swimming posture.

[0038] The above are merely specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of protection of the described technical solutions.

Claims

1. A device for monitoring fish behavior, characterized in that, The device includes: The fish body mesh construction module collects continuous underwater video stream data of the target fish, scans pixel grayscale gradient changes, identifies the contour range, extracts the geometric midline to define the spinal axis, filters feature positioning points and connects them to generate a set of triangular mesh vertices on the fish body surface. The vertex optical flow tracing module calls the vertex set of the triangular mesh on the fish surface, retrieves the same texture feature pixel region in adjacent time frames, matches the evolution trajectory of vertex spatial position, and establishes a temporal evolution vertex coordinate mapping pair. The compression potential energy calculation module constructs an affine transformation matrix from the previous frame to the real frame based on the temporal evolution vertex coordinate mapping pair, calculates the local area scaling Jacobian determinant, and performs differential operations and nonlinear exponential amplification in combination with rigid body motion reference parameters to obtain the local mesh compression potential energy index. The energy field thermodynamic mapping module maps the local grid compression potential energy index to the fish body topology, performs weighted smooth interpolation, scans the continuous energy distribution spectrum, identifies energy intensity clusters exceeding preset limits, and generates high-energy connected domains for body work. The swimming pattern identification module calculates the position of the high-energy connected domain of the body's work at the geometric center of the spinal axis, analyzes the conduction characteristics and spatial distribution pattern, matches the biomechanical motion model to determine the type, and obtains the classification results of fish tail-wagging swimming posture. The compression potential energy calculation module includes: The affine transformation construction submodule extracts the vertex coordinate groups of the same triangular unit in the previous frame and the real frame based on the temporal evolution vertex coordinate mapping pair, establishes a linear transformation equation describing the evolution of the geometric shape from the previous moment to the real moment, analyzes the equation coefficients using the least squares method, determines the affine parameters including rotation, shearing and translation components, constructs a mathematical transformation model that maps the triangular geometric shape of the previous frame to the shape of the real frame, and generates a local geometric shape transformation matrix. The area scaling calculation submodule calls the local geometric transformation matrix, extracts the diagonal elements representing the scaling ratio of the coordinate axis direction and the off-diagonal elements representing the degree of inclination between axes, performs second-order determinant algebra operation, analyzes the scaling ratio of the local mesh unit area micro-element, calculates the Jacobian determinant value representing the local area scaling state, quantifies the degree of geometric deformation of the triangular unit, and obtains the local area scaling determinant value. The potential energy calculation submodule, based on the local area scaling determinant value, introduces the rigid body motion reference parameter characterizing the non-deformation state, performs numerical difference operation, calculates the deviation between the real-time grid area change rate and the reference value, performs nonlinear exponential amplification processing on the difference result, optimizes the numerical resolution of the deformation range, and obtains the local grid compression potential energy index. The process of introducing the preset energy intensity limit reference parameter is as follows: statistically analyze the energy scalar values ​​of discrete nodes in the continuous energy distribution spectrum, calculate the arithmetic mean of the energy values ​​of the whole domain as the reference background level, calculate the root mean square deviation of the energy scalar values ​​of discrete nodes relative to the arithmetic mean, and quantify the degree of dispersion of the whole domain energy distribution. Set a confidence factor greater than 0, multiply the confidence factor by the root mean square deviation, and obtain the noise fluctuation tolerance. The arithmetic mean and the noise fluctuation tolerance are added together to output a uniquely determined preset energy intensity limit reference parameter.

2. The fish behavior monitoring device according to claim 1, characterized in that, The fish body surface triangular mesh vertex set includes node 3D coordinate data, triangular facet index table, and vertex normal vector. The temporal evolution vertex coordinate mapping pair includes inter-frame displacement correlation vector, timestamp sequence marker, and trajectory correlation coefficient. The local mesh compression potential energy index includes area scaling deformation rate, nonlinear potential energy value, and muscle contraction state identifier. The body work high-energy connected domain includes energy accumulation region contour, high-energy core distribution coordinates, and regional work intensity extrema. The fish tail-wagging swimming posture classification result includes swimming posture pattern category label, movement mechanism matching degree, and energy consumption efficiency evaluation parameters.

3. The fish behavior monitoring device according to claim 1, characterized in that, The fish body mesh construction module includes: The contour scanning and recognition submodule acquires continuous underwater video stream data of the target fish, performs pixel-level scanning operation on a single frame image matrix, calculates the gray-level gradient magnitude and direction vector between adjacent pixels, defines the boundary line between the foreground target and the background area based on the gradient change position, extracts continuous edge pixel coordinates and arranges them in a clockwise direction, identifies the geometric shape range of the organism in the image coordinate system, and generates a fish body edge contour coordinate sequence. The feature node extraction submodule calls the fish body edge contour coordinate sequence, traverses the pixels inside the contour and the boundary region, calculates the corner response function value and edge curvature extreme value of the local region, retains the pixel points whose response values ​​exceed the feature screening benchmark as key nodes, performs non-maximum suppression operation, removes collinear points and discrete noise points, and obtains the set of feature positioning points on the surface of the organism. The topology mesh generation submodule performs Delaunay triangulation based on the set of feature points on the surface of the organism, connects the three feature points with the closest spatial distance according to the empty circle property criterion, constructs non-overlapping triangular units, maximizes the minimum angle of the interior angle of the triangle, calculates the vertex index relationship and adjacency list information of the triangular units, and obtains the vertex set of the triangular mesh on the surface of the fish.

4. The fish behavior monitoring device according to claim 3, characterized in that, The vertex optical flow tracing module includes: The texture feature retrieval submodule calls the triangular mesh vertex set on the fish body surface to obtain video frame image data at the real time and the next adjacent time. It constructs a local pixel window with the mesh vertex coordinates as the center, calculates the gray-level distribution statistical features and texture gradient direction of the pixels in the window, performs cross-correlation operation or least squares difference matching within the search range of adjacent time frame images, identifies biological tissue pixel regions with the same texture gray-level features, and generates an index of texture matching regions in adjacent frames. The spatial trajectory matching submodule calculates the pixel coordinate displacement of the grid vertex in the continuous time series based on the adjacent frame texture matching region index, performs optical flow constraint operation, solves the instantaneous motion velocity vector of the vertex on the image plane, corrects the motion direction of abnormal drift points by combining smoothing constraint conditions, connects the spatial position coordinates of the same vertex at different times, matches the spatial position evolution trajectory of the vertex at different times, and obtains the vertex spatiotemporal displacement vector sequence. The temporal mapping construction submodule extracts the two-dimensional coordinate values ​​and corresponding timestamps of the vertices based on the spatiotemporal displacement vector sequence of the vertices, arranges the spatial position data of the vertices in chronological order, constructs an association index between the unique identifier of the vertex and the spatiotemporal coordinate data, identifies the motion migration relationship of discrete nodes on the surface of the organism in the time dimension, and establishes a temporal evolution vertex coordinate mapping pair.

5. The fish behavior monitoring device according to claim 1, characterized in that, The energy field thermodynamic mapping module includes: The energy topology mapping submodule calls the local mesh compression potential energy index, maps the quantized values ​​one by one to the spatial geometric nodes defined by the triangular mesh vertex set on the fish body surface, constructs a discrete scalar field characterizing the force distribution on the body surface, calculates the spatial distance weights of nodes in the local neighborhood relative to the center point, performs weighted smooth interpolation processing, fills the mesh gaps and suppresses noise fluctuations, and establishes a continuous energy distribution map. The intensity limit screening submodule scans the global grid node values ​​included in the continuous energy distribution map, introduces a preset energy intensity limit benchmark parameter, filters active nodes whose node values ​​exceed the energy intensity limit benchmark parameter, records spatial coordinates and filters background noise areas to obtain the set of pixels with excess energy. The connected component morphology extraction submodule calls the set of super-energy pixels, analyzes the neighborhood adjacency state of discrete pixels in the image space, performs morphological closing operation, bridges gaps and merges adjacent isolated point clusters, and aggregates spatially continuous pixels into independent regions according to connectivity rules to generate high-energy connected components for body work.

6. The fish behavior monitoring device according to claim 5, characterized in that, The swimming pattern recognition module includes: The spinal center localization submodule calls the high-energy connected domain of the body to extract the planar coordinate distribution data of the pixels inside the connected region, calculates the first-order geometric moment of the region, determines the position of the centroid on the image plane, introduces the linear equation parameters of the fish spinal axis, performs the projection operation from point to line, maps the centroid coordinates to the direction of the spinal axis, quantifies the distribution position of the high-energy region on the body trunk, and obtains the spinal projection coordinates of the high-energy region. The spatiotemporal feature analysis submodule, based on the spinal projection coordinates of the high-energy region, arranges the projection position data in consecutive frames in a time sequence, calculates the instantaneous velocity vector and acceleration change rate of the coordinate point moving along the spinal direction, analyzes the phase delay relationship transmitted from the head to the tail, calculates the eccentricity and compactness parameters in combination with the spatial distribution range of the connected domain, analyzes the spatiotemporal evolution law of the work area of ​​the organism, and generates the spatiotemporal conduction feature vector of energy flow. The motion model matching submodule, based on the energy flow spatiotemporal conduction feature vector, calls the feature parameter set of standard swimming strokes in the preset biomechanical motion model library, performs similarity calculation or Euclidean distance measurement operation in the vector space, finds the benchmark model category that matches the real-time feature vector, determines the motion mechanism type of the target object in the real-time observation period, and obtains the classification result of fish tail-wagging swimming stroke.

7. The fish behavior monitoring device according to claim 6, characterized in that, The process of calculating the instantaneous velocity vector and acceleration rate of change of the coordinate point moving along the spine direction is as follows: extract the high-energy region spine projection coordinates corresponding to the real-time sampling frame and the previous sampling frame, calculate the algebraic difference between the coordinate values ​​of the two time points in the direction of the spine axis, and obtain the displacement increment. Obtain the frame rate parameters of the video acquisition device and calculate the inter-frame time interval. Divide the displacement increment by the inter-frame time interval to obtain the instantaneous velocity vector characterizing the first derivative. Extract instantaneous velocity vector data from three consecutive time points, calculate the change in the difference between adjacent velocity vectors, divide the change in the difference by the inter-frame time interval, and obtain the acceleration value representing the second derivative.

8. A method for monitoring fish behavior, characterized in that, The fish behavior monitoring device according to any one of claims 1-7 comprises the following steps: S1: Collect continuous underwater video stream data of the target fish, scan the pixel grayscale gradient changes within the video frame, identify the fish body outline range, filter the feature positioning points within the outline and connect them to generate a set of triangular mesh vertices on the fish body surface. S2: Call the triangular mesh vertex set on the surface of the fish, retrieve pixel regions with the same texture features in adjacent time frames, match the evolution trajectory of vertices in spatial position, and establish temporal evolution vertex coordinate mapping pairs; S3: Based on the temporal evolution vertex coordinate mapping pair, construct the morphological affine transformation matrix from the previous frame to the real frame, use the matrix elements to calculate the local area scaling Jacobian determinant, combine the rigid body motion reference parameters to perform differential operation on the Jacobian determinant, and perform nonlinear exponential amplification processing on the differential operation value to obtain the local mesh compression potential energy index. S4: Map the local mesh compression potential energy index to the fish body topology, perform weighted smooth interpolation, scan the continuous energy distribution spectrum, identify the clustered regions where the energy intensity exceeds the preset limit, and generate the high-energy connected domain for body work. S5: Calculate the position of the high-energy connected domain of the body doing work at the geometric center of the spinal axis, analyze the conduction characteristics and spatial distribution of the connected domain, match the biomechanical motion model parameters to determine the type, and obtain the classification results of fish tail-wagging swimming posture.

Images