River crab growth state monitoring and evaluation system and method

Abstract

This invention relates to the field of river crab growth status monitoring technology, specifically to a river crab growth status monitoring and evaluation system and method. In this invention, the target to be monitored is located and a background microfluidic field vector reflecting environmental stimuli is constructed. Fluid dynamics analysis is then performed by combining the extracted target centroid point with appendage joint points. The macroscopic drift cross-correlation characteristic representing the overall motion response and the joint disturbance resistance damping characteristics reflecting limb resistance are calculated and input into a manifold-coupled visual dynamics model for discrimination. This method effectively utilizes the interaction characteristics between fluid and target, overcomes morphological similarity interference, and accurately identifies the biological state of live river crabs, molted river crabs, and dead river crabs in dynamic water bodies. Based on this, scientific growth monitoring and evaluation data are generated, improving the accuracy and intelligence level of aquaculture monitoring.

Description

Technical Field

[0001] This invention relates to the field of river crab growth status monitoring technology, specifically to a river crab growth status monitoring and evaluation system and method. Background Technology

[0002] With the continuous advancement of smart aquaculture technology, non-contact monitoring methods based on underwater computer vision have been gradually applied to the farming management of river crabs, aiming to achieve precise feeding and growth assessment through real-time perception of farmed organisms. However, in actual underwater unstructured scenarios, river crabs have the biological characteristic of periodic molting. Their molted shells have a very high similarity to live and dead river crabs in terms of morphological texture, color characteristics, and geometric contours. Conventional deep learning networks based on appearance feature extraction are often insufficient to effectively distinguish the three at the static image level. Adding to the complexity, the aquaculture water is usually in a continuous, non-static state. Affected by aeration equipment, bottom currents, or disturbances from surrounding organisms, a dynamic microflow field is prevalent in the water. This environmental flow field, as an external excitation force, drives the less dense and lighter hollow molted shells or dead river crabs that have lost neuromuscular control to passively drift with the current, and their appendages to sway with the water flow. Such passive displacement and micro-movements are easily misjudged by conventional moving target detection algorithms as active crawling or feeding behavior of living organisms in the temporal visual domain.

[0003] Existing monitoring technologies mostly focus on static appearance recognition or simple moving foreground segmentation, lacking in-depth analysis of the coupling relationship between underwater environmental flow dynamics factors and the mechanical motion response of target objects. It is difficult to effectively perceive the anti-dampening characteristics of targets through visual means under dynamic water flow interference, resulting in large statistical biases in the assessment data of live crab stock, mortality rate and molting rate, which is difficult to meet the actual needs of high-precision automated aquaculture decision-making. Summary of the Invention

[0004] The purpose of this invention is to provide a system and method for monitoring and evaluating the growth status of river crabs, in order to solve the problems mentioned in the background art. Specific technical problems include how to utilize background microfluidic field vectors as environmental excitation and analyze the hydrodynamic response of the centroid and appendage joints of the monitored target to this flow field excitation, as well as the joint's disturbance rejection and damping characteristics. This addresses the technical problem of inaccurate growth status monitoring and evaluation data caused by the difficulty in accurately distinguishing between live river crabs, molted river crabs, and dead river crabs due to their similar morphology and the influence of water flow disturbances in dynamic aquatic environments.

[0005] To achieve the above objectives, one of the objectives of this invention is a method for monitoring and evaluating the growth status of river crabs, comprising the following steps:

[0006] S1. Acquire sequential images of the aquaculture area and locate the target to be monitored in the sequential images, specifically including:

[0007] The underwater optical camera continuously collects primary analog video signals of the underwater environment using a fixed bracket installed at the bottom of the aquaculture area, and converts the primary analog video signals into digital sequence images using a video capture card;

[0008] Pixel-level enhancement processing is performed on digital sequence images to generate enhanced digital sequence images; the pixel-level enhancement process uses a dark channel prior image dehazing algorithm combined with a contrast-limited adaptive histogram equalization algorithm.

[0009] A convolutional neural network model containing a deep residual operation module and a multi-scale feature fusion structure is constructed. The enhanced digital sequence image is input into the convolutional neural network model, and the bounding box coordinates of the target to be monitored in the current frame image and the target confidence score are output.

[0010] The Kalman filter multi-target tracking algorithm is adopted. Based on the intersection-union ratio of the bounding box coordinates between adjacent frames and the association cost matrix, the target to be monitored in the digital sequence image is associated across frames and assigned a unique identification code to complete the continuous localization of the target to be monitored.

[0011] Step S1 effectively improves the image degradation problem caused by underwater turbidity by combining underwater optical imaging devices with enhancement algorithms, thereby enhancing the clarity of visual information. By using a convolutional neural network model that includes a depth residual calculation module, the target to be monitored can be identified and located more accurately in complex backgrounds. At the same time, by combining the Kalman filter multi-target tracking algorithm, continuous and stable trajectory association of the target to be monitored in time-series images is achieved, providing a reliable target position basis for subsequent dynamic analysis.

[0012] S2. Extract the motion trajectories of suspended particles in the background region of the sequence images and construct a temporal background microfluidic field vector, specifically including:

[0013] Based on the bounding box coordinates of each target to be monitored in the current frame, a target inverse mask is generated in the enhanced digital sequence image, and the image region in the enhanced digital sequence image excluding the area covered by the target inverse mask is defined as the background region of interest.

[0014] The Gaussian mixture background modeling algorithm is used to segment the background region of interest, extract dynamically changing pixel groups, and determine the dynamically changing pixel groups whose pixel area is within a preset threshold as floating particle feature points.

[0015] The pyramid Lucas-Carnard sparse optical flow algorithm is used to calculate the pixel displacement vector of each suspended particle feature point between the current frame and the next frame.

[0016] The effective pixel displacement vectors after removing outliers are calculated by vector weighted average to obtain the background average flow velocity vector and background average flow direction vector of the current frame.

[0017] The background average velocity vector and background average flow direction vector calculated for each frame in the continuous time series are combined in time stamp order to construct a temporal background microflow field vector.

[0018] Step S2 effectively captures the dynamic changes of water bodies by eliminating the target region and locking the background region of interest, using suspended particles as natural tracers. The displacement of the particles is calculated by the Pyramid Lucas-Cannard sparse optical flow algorithm, and the fluid dynamic state of the water environment is quantified by weighted averaging. This constructs a background microflow field vector that reflects the environmental excitation effect, which provides key environmental reference data for analyzing the passive force situation of the target object under the action of water flow.

[0019] S3. Extract the key skeletal points of the target to be monitored. The key skeletal points include the target's centroid and appendage joints, specifically including:

[0020] Based on the bounding box coordinates of each target to be monitored in the current frame, the region of interest image of the target to be monitored is extracted from the enhanced digital sequence image;

[0021] The image of the region of interest of the target to be monitored is input into a cascaded pyramid pose estimation neural network model that has been pre-trained and converged based on a dataset of key points labeled on the skeleton of the crab.

[0022] The cascaded pyramid-shaped attitude estimation neural network model outputs a two-dimensional Gaussian response heatmap corresponding to multiple channels of the anatomical structure of the target to be monitored;

[0023] For each channel's two-dimensional Gaussian response heatmap, a maximum coordinate search is performed to extract the pixel position with the highest response value as the key point coordinate. The pixel position corresponding to the geometric center of the crab's cephalothorax is marked as the target centroid point, and the pixel position corresponding to the joint connection of the crab's two walking legs and chelipeds is marked as the appendage joint point, thus completing the extraction of the skeleton key points of the target to be monitored.

[0024] Step S3 utilizes a cascaded pyramid-shaped pose estimation neural network model to delve into the fine anatomical structure of the target, effectively locating the target's centroid and limb joints under adversarial occlusion or complex poses. This fine-grained feature extraction not only determines the overall position of the target but also captures the spatial configuration of the limbs, providing necessary morphological and geometric support for distinguishing the active control behavior of living organisms from the passive mechanical homing of non-living targets.

[0025] S4. Based on the background micro-flow field vector, perform hydrodynamic response analysis on the target to be monitored, calculate the macroscopic drift cross-correlation between the displacement vector of the target centroid and the background micro-flow field vector, specifically including:

[0026] Based on the pixel coordinates of the target centroid point in the continuous time series, calculate the displacement vector of the target centroid point in the current frame relative to the previous frame, and define it as the target centroid point displacement vector.

[0027] Call the background average velocity vector and background average flow direction vector corresponding to the current frame from the background microflow field vector;

[0028] Calculate the cosine of the angle between the target centroid displacement vector and the background average flow direction vector, and use it as the flow direction consistency coefficient.

[0029] The ratio of the magnitude of the displacement vector of the target centroid point to the magnitude of the average velocity vector of the background is calculated and used as the velocity response ratio coefficient.

[0030] A linear weighted summation algorithm is used to fuse the flow direction consistency coefficient and the flow velocity response ratio coefficient to calculate the macroscopic drift cross-correlation that characterizes the degree of drift of the monitored target with the flow.

[0031] Step S4 can couple the macroscopic motion state of the target with the environmental water flow state. By calculating the flow direction consistency coefficient and the flow velocity response ratio coefficient, the degree of the target drifting with the current is quantitatively assessed. The obtained macroscopic drift cross-correlation index effectively reflects the interaction between the mass of the target object and the thrust of the water flow, which helps to distinguish live crabs with the ability to grip the ground from lighter crab molts or dead individuals that are prone to drifting from the perspective of motion mechanism.

[0032] S5. Track the temporal micro-motion trajectory of the appendicular joints relative to the trunk connection point, and calculate the joint disturbance rejection damping characteristics of the appendicular joints by combining the excitation effect of the background micro-flow field vector, specifically including:

[0033] Define the target centroid point of the current frame as the origin of the coordinate system, and calculate the relative coordinate vector of each appendage joint point with respect to the target centroid point in the current frame.

[0034] Arrange the relative coordinate vectors of all frames in the continuous time series in the order of timestamps to construct the temporal micro-motion trajectory of the limb joint point relative to the trunk connection point;

[0035] The standard deviation of the time-domain micro-motion trajectory within a preset time window is calculated and defined as the micro-motion response intensity of the appendage.

[0036] The background average velocity vector corresponding to the current frame in the background micro-flow field vector is acquired synchronously, and its modulus is calculated as the fluid excitation intensity.

[0037] The ratio of fluid excitation intensity to limb micromotion response intensity is calculated and determined as the joint disturbance resistance damping characteristic of the limb joint.

[0038] Step S5 quantifies the response and resistance characteristics of the target appendage structure to water flow impact by tracing the temporal micro-motion trajectory of the appendage joints relative to the target centroid and combining it with the fluid excitation intensity of the background micro-flow field vector. The calculated joint disturbance resistance damping characteristics reflect the energy dissipation mechanism of the target at the dynamic level, which helps to identify the active adjustment ability and limb structure stability of live crabs. At the same time, it can reflect the passive following and highly compliant wave motion characteristics exhibited by crab molts or dead crabs under water flow disturbance.

[0039] S6. Input the macroscopic drift cross-correlation and joint disturbance rejection damping characteristics into the manifold-coupled visual dynamics model, and output the biological state of the target to be monitored. The biological state includes live crabs, crab molts, and dead crabs, specifically including:

[0040] The macroscopic drift cross-correlation and joint disturbance rejection damping characteristics are combined to construct a two-dimensional state feature vector, which is then normalized.

[0041] The processed two-dimensional state feature vector is input into a convergent manifold-coupled visual dynamics model pre-trained based on bio-tagged sample data.

[0042] The radial basis function integrated within the manifold coupled visual dynamics model is used to map the two-dimensional state feature vector to a high-dimensional feature space, and the geometric margin between the mapped feature vector and the preset optimal classification hyperplane is calculated.

[0043] The category is determined based on the value of the geometric interval and its location within the decision area, thus identifying the biological status of the target to be monitored.

[0044] Step S6 utilizes the manifold-coupled visual dynamics model and its internal radial basis kernel function to map the physical-level dynamic characteristics (i.e., macroscopic drift cross-correlation and joint disturbance rejection damping characteristics) to a high-dimensional space, enhancing the separability of features. By comprehensively analyzing the overall drift degree of the target and the local limb damping characteristics, this manifold-coupled visual dynamics model can accurately determine the three states of live crabs, crab molts, and dead crabs within the decision-making area, effectively reducing the misjudgment rate caused by visual similarity.

[0045] S7. Statistically determine the target number and morphological parameters of live river crabs, and generate growth monitoring and evaluation data, specifically including:

[0046] Targets that are determined to be live river crabs are selected, and the real-time inventory of live river crabs is counted.

[0047] For each live crab, the bounding box coordinates of its current frame are called to calculate the shell morphology features, and the shell morphology features are converted into physical size data by combining the preset camera imaging calibration coefficients.

[0048] Using a pre-constructed allometric power function model of the relationship between carapace physical size and biomass, the estimated weight of each live crab was calculated based on physical size data.

[0049] The average value and distribution statistics of the real-time stock, physical size data of all live crabs, and estimated weight of each individual are calculated to generate a structured report containing population density, average size parameters, and a histogram of population weight distribution, which is defined as growth monitoring and evaluation data.

[0050] Step S7 transforms the visual recognition results into specific quantitative indicators for aquaculture. It uses an allometric growth power function model to calculate the biophysical weight from the image pixel size. The generated growth monitoring and evaluation data covers multiple dimensions such as population density, size parameters, and weight distribution, providing aquaculture personnel with intuitive and digital aquaculture status reports, which helps to scientifically assess the growth status and stocking conditions of crabs.

[0051] The second objective of this invention is to provide a system for monitoring and evaluating the growth status of river crabs, comprising an image acquisition and target localization module, a background microfluidic field construction module, a skeleton key point extraction module, a macroscopic drift cross-correlation analysis module, a joint disturbance resistance damping characteristic calculation module, a manifold coupled visual dynamics recognition module, and a growth monitoring and statistical evaluation module, wherein:

[0052] The image acquisition and target localization module acquires a sequence of images of the aquaculture area and locates the target to be monitored in the sequence of images;

[0053] The background microflow field construction module extracts the motion trajectories of suspended particles in the background region of sequential images and constructs a temporal background microflow field vector.

[0054] The skeleton key point extraction module extracts the skeleton key points of the target to be monitored, including the target centroid point and appendage joint points;

[0055] The macroscopic drift cross-correlation analysis module performs fluid dynamic response analysis on the target under monitoring based on the background micro-flow field vector, and calculates the macroscopic drift cross-correlation between the displacement vector of the target centroid point and the background micro-flow field vector.

[0056] The joint disturbance resistance and damping characteristic calculation module tracks the temporal micro-motion trajectory of the appendiceal joint points relative to the trunk connection points, and calculates the joint disturbance resistance and damping characteristics of the appendiceal joint points by combining the excitation effect of the background micro-flow field vector.

[0057] The manifold-coupled visual dynamics recognition module inputs the macroscopic drift cross-correlation and joint disturbance rejection damping characteristics into the manifold-coupled visual dynamics model and outputs the biological state of the target to be monitored, including live crabs, crab molts, and dead crabs.

[0058] The growth monitoring and statistical evaluation module statistically determines the target number and morphological parameters of live river crabs and generates growth monitoring and evaluation data.

[0059] Compared with the prior art, the beneficial effects of the present invention are:

[0060] This invention quantifies the dynamic fluid excitation effect of the aquaculture water environment by constructing a background microfluidic field vector, and introduces a fluid dynamics response analysis mechanism by extracting the target centroid and appendage joints. By calculating the macroscopic drift cross-correlation and joint disturbance resistance damping characteristics, it can deeply analyze the interaction characteristics between the monitored target and the aquatic environment from the perspective of physical motion mechanism, effectively overcoming the technical problem of traditional visual methods in accurately distinguishing live crabs, crab molts, and dead crabs in dynamic underwater scenes due to their highly similar appearance. The invention utilizes a manifold coupled visual dynamics model to perform high-dimensional mapping and classification of the above dynamic characteristics, significantly improving the robustness of target organism status recognition. The growth monitoring and evaluation data generated on this basis covers key indicators such as population density and estimated weight, realizing non-contact perception of the growth status of underwater aquaculture organisms, and providing scientific and reliable data support for precision aquaculture management. Attached Figure Description

[0061] Figure 1 This is a schematic diagram of the overall method steps of the present invention;

[0062] Figure 2 This is a schematic diagram of the core process from steps S2 to S5 of the present invention;

[0063] Figure 3 This is a schematic diagram of the core process of steps S6 to S7 of the present invention;

[0064] Figure 4 This is a schematic diagram illustrating the distribution characteristics of the macroscopic drift cross-correlation and joint disturbance rejection damping characteristics of the present invention.

[0065] Figure 5 This is a schematic diagram of the overall module flow of the present invention.

[0066] In the figure: 100, Image acquisition and target localization module; 200, Background microfluidic field construction module; 300, Skeleton key point extraction module; 400, Macroscopic drift cross-correlation analysis module; 500, Joint disturbance resistance damping characteristic calculation module; 600, Manifold coupling visual dynamics recognition module; 700, Growth monitoring and statistical evaluation module. Detailed Implementation

[0067] The technical solutions in 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.

[0068] Next, please refer to Figure 1 One of the objectives of this embodiment is to provide a method for monitoring and evaluating the growth status of river crabs.

[0069] The specific steps are as follows:

[0070] S1. Acquire a sequence of images of the aquaculture area and locate the target to be monitored in the sequence of images, specifically including:

[0071] An underwater optical camera device installed on a fixed bracket at the bottom of the aquaculture area continuously collects primary analog video signals of the underwater environment, and then converts the primary analog video signals into digital sequence images via a video capture card. ;

[0072] A dark channel prior image dehazing algorithm combined with a contrast-limited adaptive histogram equalization algorithm is used to process digital sequence images. Perform pixel-level enhancement processing to generate enhanced digital sequence images. Specifically, the first step is to perform dehazing based on dark channel prior theory, a process that iterates through the digital image sequence. For each pixel in the image, the brightness value in the red, green, and blue color channels is analyzed. The minimum brightness value within a local neighborhood is searched to construct a dark channel map. Based on the distribution characteristics of non-zero pixels in the dark channel map, the transmittance distribution of the water medium and the intensity of background scattered light are estimated. Finally, the inverse transform formula of the atmospheric scattering physical model is used to peel away the digital image sequence. The water body scattered light component in the image was used to reconstruct the scene's irradiance information; subsequently, to further improve the dehazing digital sequence images... To improve the discernibility of details, adaptive histogram equalization with limited contrast is implemented for digital image sequences. The image is divided into several non-overlapping rectangular sub-regions. A gray-level histogram is calculated for each sub-region. High-frequency peaks in the gray-level histogram are truncated using a preset amplitude limiting threshold to suppress excessive amplification of background noise. The pixels of the truncated portion are then evenly distributed across the other gray levels of the gray-level histogram to reshape the cumulative distribution function. This process generates an enhanced digital sequence image with high definition, high contrast, and uniform brightness from the original digital sequence image. ;

[0073] Construct a convolutional neural network model that includes a deep residual operation module and a multi-scale feature fusion structure; and enhance the digital sequence image. The input data is uniformly adjusted to a fixed pixel size. Deep semantic features are extracted sequentially through stacked residual operation modules within the convolutional neural network model. Lateral connection structures are then used to fuse deep and shallow features to enhance the ability to capture features of small targets. The fused feature map is then used to generate candidate anchor boxes via a region suggestion subnetwork. After undergoing region feature calibration (correcting the spatial correspondence between the feature map and the original image region), it enters the fully connected layer, outputting the current frame time. The image of the first Coordinates of the bounding box of each target to be monitored (in The coordinates of the top left pixel. These are the width and height of the rectangular bounding box, respectively, and the target confidence score. The convolutional neural network model is trained on a standardized sample dataset containing pre-set true bounding box labels. A joint loss function is defined, which includes classification cross-entropy loss and bounding box position regression loss. The model parameters are iteratively optimized through the stochastic gradient descent algorithm until convergence.

[0074] A Kalman filter multi-target tracking algorithm is used, based on the intersection-union ratio (IU) of the bounding box coordinates between adjacent frames. and the associated cost matrix Cross-frame trajectory association is performed on the target to be monitored in digital image sequences, and a unique identification code is assigned. This enables continuous positioning of the monitored targets; among which ,in A function that calculates the pixel area of ​​a geometric region; Indicates the current frame time The index of the candidate prediction box in the image; The above formula is used to quantize the previous frame time. With the current frame time In the image, the spatial overlap between different predicted boxes is considered; by minimizing the association cost matrix, the identity matching problem of multiple targets in continuous time series is solved, ensuring that subsequent steps are processed for the same physical target; the Kalman filter multi-target tracking algorithm aims to solve the trajectory interruption problem of the monitored target due to occlusion or motion blur in dynamic scenes, specifically including:

[0075] It establishes a system state vector containing the center coordinates of the target to be monitored, the area scale of the bounding box, the aspect ratio, and the rates of change of the above four variables in the horizontal and vertical directions. In each step of the recursive loop, the time update stage is executed first. The linear uniform motion model is used to deduce the state prior estimate and prior error covariance matrix of the current frame based on the optimal posterior estimate of the previous frame. Then, the measurement update stage is entered. The target bounding box observation data of the current frame output by the positioning step is used as the measurement value. The residual between the measurement and the state prior estimate is calculated, and the Kalman gain matrix is ​​solved by combining the statistical characteristics of system process noise and measurement noise. Finally, the state prior estimate is weighted and corrected using the Kalman gain matrix to obtain the optimal posterior state estimate of the current frame and update the error covariance matrix. In this way, the smooth prediction and position correction of the motion trajectory of the same target to be monitored in a continuous time series are achieved.

[0076] Please see Figure 2 S2. Extract the motion trajectories of suspended particles in the background region of the sequence image and construct a temporal background microfluidic field vector, specifically including:

[0077] Based on the current frame time output in step S1 The coordinates of the bounding box of each target to be monitored In the enhanced digital sequence image Generating a reverse mask for the target And enhance the digital sequence image. Remove the target reverse mask The image region outside the covered area is defined as the background region of interest. ;

[0078] Using Gaussian mixture background modeling algorithm to identify regions of interest in the background Foreground segmentation is performed to extract dynamically changing pixel groups, and the pixel area is set within a preset threshold. The dynamically changing pixel clusters within the area were identified as suspended particle feature points. ,in These represent the lower and upper threshold values ​​for the pixel area of ​​suspended particles, respectively, typically ranging from [5, 50] pixels. These values ​​are determined based on the camera resolution and the average size of suspended matter in the water to filter out noise. The Gaussian mixture background modeling algorithm targets the background region of interest. The brightness variation of each pixel location over a continuous time series is analyzed, and an adaptive mixing model composed of multiple weighted normal distribution functions is established to fit the color probability density distribution of the water background. When processing each new frame, the matching degree between the current pixel brightness value and each normal distribution function in the mixing model is calculated. Based on the matching results, the mean, variance, and normalized weight coefficients of each distribution function are updated in real time. If the current pixel brightness value cannot match any existing normal distribution function, the distribution function with the lowest weight coefficient is replaced to adapt to the gradual change in illumination. The distribution functions are sorted according to the ratio of their weight coefficients to their variance parameters. The set of distribution functions that ranks highest and whose combined weight coefficients reach a predetermined proportion is determined as a stable water background model, while pixels that do not match this model are determined as foreground targets, thus being selected from the background region of interest. It accurately segments dynamically changing groups of pixels.

[0079] The pyramid Lucas-Cardard sparse optical flow algorithm is used to calculate the feature points of each suspended particle (index is ). At the current frame time With the next frame pixel displacement vector between The Pyramid Lucas-Carnard sparse optical flow algorithm, based on the assumption of constant image grayscale and local optical flow smoothing constraints, aims to solve the problem of suspended particle feature points in the current frame. With the next frame This paper addresses the problem of tracking large-span displacements caused by rapid flow. First, it performs multi-level downsampling on digital image sequences, constructing an image pyramid structure with increasing resolution. An iterative computation strategy, progressing from coarse to fine, is employed. Initial optical flow vectors are calculated at the top level (low-resolution image layer) of the pyramid, transforming large-span displacements into small displacements that satisfy the linearization assumption. The optical flow vectors calculated at the previous level are then passed as initial estimates to the next level (high-resolution image layer). At this level, spatial gradient information within a local window is used to calculate the residual displacement using the least squares method to correct the initial estimate. This process is repeated layer by layer down to the original resolution image layer. Finally, by superimposing the corrected displacements from each pyramid level, the pixel displacement vector of each suspended particle feature point is calculated. ;

[0080] After removing outliers from the pixel displacement vectors using the Euclidean distance formula, the remaining... The vector weighted average of the effective pixel displacement vectors is calculated to obtain the current frame time. Background average velocity vector Average flow vector with background ,Right now ,in This represents the total number of valid suspended particle feature points remaining after removing outliers; it is a natural number. Indicates the first The confidence weight coefficient of each suspended particle displacement vector has a value range of (0,1] and is determined based on the product correlation value of optical flow matching. ;

[0081] The background average velocity vector and background average flow direction vector calculated for each frame in the continuous time series are combined in time stamp order to construct a temporal background microflow field vector. ,Right now ,in This represents the total length (number of frames) of the time series used to construct the background microfluidic field vector, typically ranging from [30, 300] frames.

[0082] Step S2 filters out individual disturbance errors by statistically analyzing the microscopic motion trends of suspended particle groups within the region of interest in the background, constructing a model that accurately reflects the water body at the current frame time. The temporal background microflow field vectors of the macroscopic flow state (flow velocity and flow direction) provide an environmental reference for subsequent analysis of the target force.

[0083] S3. Extract the key skeletal points of the target to be monitored. The key skeletal points include the target's centroid and appendage joints, specifically including:

[0084] Based on the current frame time output in step S1 The coordinates of the bounding box of each target to be monitored From the enhanced digital sequence image Extracting the region of interest from the image of the target to be monitored and the image of the region of interest of the target to be monitored. Input a pre-trained, convergent cascaded pyramid pose estimation neural network model based on a crab skeleton keypoint annotation dataset; construct the cascaded pyramid pose estimation neural network model and input the region of interest image of the target to be monitored. Pixel values ​​are normalized and used as input. The input is processed through a multi-level, interconnected hourglass-shaped feature extraction module within a cascaded pyramid-style pose estimation neural network model. Each module consists of a downsampling layer capturing global context information and an upsampling layer restoring spatial resolution. Multi-scale features are fused between layers via skip connections. The final output layer generates a feature vector using convolution operations. ( The two-dimensional Gaussian response heatmap of the cascaded pyramid-shaped attitude estimation neural network model represents the total number of feature channels output by the model (its value is equal to the total number of keypoints of the target being monitored). ,in and These represent the x and y coordinates of the two-dimensional Gaussian response heatmap in the pixel coordinate system, respectively; 2D Gaussian response heatmap The pixel brightness value represents the probability that a specific key point exists at that location; the model is trained based on a dedicated dataset containing the ground truth coordinates of preset key points (centroid and joints), and the ground truth coordinates are converted into a two-dimensional Gaussian distribution heatmap as a supervision label. The mean squared error loss function is used to minimize the pixel-level difference between the predicted heatmap and the labeled heatmap.

[0085] The output of the cascaded pyramid pose estimation neural network model corresponds to the anatomical structure of the target being monitored. Two-dimensional Gaussian response heatmap of each channel For each channel's 2D Gaussian response heatmap, a maximum coordinate search is performed to extract the pixel location with the highest response value as the keypoint coordinates. Mark the pixel position corresponding to the geometric center of the cephalothorax of the crab as the target centroid point. The pixel positions corresponding to the joints of the walking legs and chelipeds on both sides of the crab are marked as appendage joint points. This completes the framework of key points for the monitored targets. Extract; among which ; ,in The total number of appendage joints is determined based on the anatomical structure of the crab (e.g., 8 walking joints plus 2 cheliped joints). The above formula aims to convert the two-dimensional Gaussian response heatmap output by the cascaded pyramidal attitude estimation neural network model into precise geometric pixel coordinates, thereby deconstructing the anatomical structure of the target to be monitored at the pixel scale and providing basic geometric data for subsequent kinematic and dynamic analysis.

[0086] S4. Based on the background micro-flow field vector, perform hydrodynamic response analysis on the target to be monitored, calculate the macroscopic drift cross-correlation between the displacement vector of the target centroid and the background micro-flow field vector, specifically including:

[0087] Based on the centroid point of the target to be monitored in the continuous time series obtained in step S3 The pixel coordinates are used to calculate the centroid of the target using the vector difference method. At the current frame time Compared to the previous frame The displacement vector is defined as the displacement vector of the target centroid point. ,Right now ;

[0088] The temporal background microfluidic field vector constructed in step S2 is invoked. The corresponding time in the current frame Background average velocity vector Average flow vector with background ; Calculate the displacement vector of the target centroid using the dot product formula. Average flow vector with background The cosine of the angle between them is used as the flow direction consistency coefficient. ,Right now ;

[0089] Calculate the displacement vector of the target centroid point Modulus length and background average velocity vector The ratio of the modulus to the velocity response ratio is used as the velocity response ratio coefficient. ,Right now ;

[0090] A linear weighted summation algorithm is used (weight factor is ). By fusing the flow direction consistency coefficient and the flow velocity response ratio coefficient, a macroscopic drift cross-correlation coefficient is calculated to characterize the degree of drift of the monitored target with the flow. ,Right now ,in and These represent the weighting factors for flow direction consistency and flow velocity response ratio, respectively, both ranging from (0,1) and satisfying the following conditions: The value is determined based on the contribution rate of flow velocity and flow direction to the drift determination in the experiment;

[0091] Step S4 quantifies the coupling degree between the target centroid displacement vector and the background environment flow field vector in terms of direction and velocity, numerically characterizing whether the target under monitoring is in a passive drifting state of "following the flow". The larger the value, the more significant the passive drift characteristic.

[0092] S5. Track the temporal micro-motion trajectory of the appendicular joints relative to the trunk connection point, and calculate the joint disturbance rejection damping characteristics of the appendicular joints by combining the excitation effect of the background micro-flow field vector, specifically including:

[0093] The current frame time output in step S3 Target centroid Defined as the origin of the coordinate system, calculate each appendage joint point output in step S3. At the current frame time Enhanced digital sequence images The relative coordinate vector of the target centroid point ,Right now ; and the relative coordinate vectors of all frames in the continuous time series Arranged in time stamp order, the temporal micro-motion trajectory of the limb joint points relative to the trunk connection points is constructed;

[0094] Calculate the time-domain micro-motion trajectory within a preset time window The standard deviation within the range is defined as the intensity of the micromotion response of the appendage. ;Right now ,in This represents the preset time window used to calculate the standard deviation of the micro-motion trajectory. The unit is frames, which are determined based on the video frame rate and the micro-motion frequency characteristics of the crab. The value is usually 25-50 frames. Indicates the sliding time window Discrete-time variable (frame index) during traversal and summation; Indicates within the preset time window Internal, relative coordinate vector of limb joint points The arithmetic mean.

[0095] Synchronously acquire the temporal background microfluidic field vector constructed in step S2 Current frame time Corresponding background average velocity vector And calculate its modulus as the fluid excitation intensity. Based on the principle of mechanical vibration, the ratio of fluid excitation intensity to appendage fretting response intensity is calculated using division operations. This ratio is used to quantitatively characterize the mechanical stiffness retention capability of the appendages of the monitored target under water flow impact, and is determined as the joint disturbance rejection damping characteristic of the appendage joints. ;Right now ;

[0096] Step S5 utilizes the definition of stiffness in physics (the ratio of force to deformation), treating the background flow field as an excitation force and the micro-motion of the appendage joints relative to the torso as a deformation response, and then calculates... Quantify the mechanical stiffness retention capability of the limb under the impact of water flow (i.e., disturbance resistance damping characteristics).

[0097] Please see Figure 3 S6. Input the macroscopic drift cross-correlation and joint disturbance rejection damping characteristics into the manifold-coupled visual dynamics model, and output the biological state of the target to be monitored. The biological state includes live crabs, crab molts, and dead crabs, specifically including:

[0098] The macroscopic drift cross-correlation calculated in step S4 The joint disturbance rejection damping characteristics calculated in step S5 Combined to construct a two-dimensional state feature vector The two-dimensional state feature vector is then dimensionless using the maximum-minimum normalization formula. ,Right now ,in and These represent the minimum and maximum values ​​of the two-dimensional state feature vectors obtained statistically from the training sample set, respectively, and are used for normalization processing; This represents the matrix transpose operation;

[0099] A manifold-coupled visual dynamics model is constructed (a nonlinear classifier based on the principle of structural risk minimization). The model construction process involves retrieving samples containing deterministic state labels (such as live, molted, and dead) and their corresponding two-dimensional state feature vectors from a historical database. The training set is constructed; during the training phase, the manifold-coupled visual dynamics model automatically finds the optimal Lagrange multipliers by solving a convex quadratic programming problem. Bias term of the classification decision hyperplane To maximize the geometric margin of different class samples in the mapped high-dimensional feature space; the manifold-coupled visual dynamics model employs a radial basis kernel function. When processing nonlinearly separable data, its width parameter The optimal selection is made in the training set through gridded search and cross-validation to ensure the model's generalization ability.

[0100] The processed two-dimensional state feature vector The input is a convergent manifold-coupled visual dynamics model pre-trained based on biotagged sample data. The radial basis function integrated within the manifold-coupled visual dynamics model is then utilized. (in (For support vector centers) The two-dimensional state feature vectors are mapped to a high-dimensional feature space; in the high-dimensional feature space, the geometric margin between the mapped feature vectors and the pre-defined optimal classification hyperplane is calculated. ,Right now ,in This represents the total number of support vectors determined after training the manifold-coupled visual dynamics model, which is automatically determined by the training algorithm. Indicates the first The Lagrange multipliers corresponding to each support vector are obtained by the training and optimization process of the manifold coupled visual dynamics model. Indicates the first Each support vector corresponds to a category label value (e.g., +1 or -1). The width parameter of the radial basis function is greater than 0 and is determined by cross-validation or grid search optimization. The bias term (intercept) representing the classification decision hyperplane is obtained by training the manifold-coupled visual dynamics model.

[0101] Based on the value of the geometric interval and its location in the decision region Determine the category; if the determination result falls into the first decision area. Output live river crabs when the judgment result falls into the second decision area. Output the crab molt when the judgment result falls into the third decision area. The system outputs the status of dead crabs in real time, thereby determining the biological state of the target being monitored. ,Right now ,in , and Representing biological states respectively Enumerated category labels for live river crabs, river crab molts, and dead river crabs;

[0102] Step S6 utilizes the principle of support vector machines to construct a nonlinear decision boundary in a high-dimensional manifold space. By integrating macroscopic drift cross-correlation and microscopic joint anti-disturbance damping characteristics, it achieves accurate classification of different biological states (living, molting, and dead) under similar visual appearances.

[0103] S7. Statistically determine the target number and morphological parameters of live river crabs, and generate growth monitoring and evaluation data, specifically including:

[0104] Iterate through the biological state of each target to be monitored output in step S6. The target objects that are determined to be live river crabs are screened out, and the real-time inventory of live river crabs is counted using a cumulative counter. ,Right now ,in This represents an indicator function, which takes the value 1 when the condition within the parentheses is true (i.e., the organism is alive), and takes the value 0 otherwise.

[0105] For each live crab, call the current frame time generated in step S1. coordinates of the bounding box ,in The morphological features of the carapace are determined by calculating the geometric length, width, and pixel span of the rectangular bounding box. And combined with preset camera imaging calibration coefficients Convert the morphological characteristics of the carapace into physical size data. ,Right now ,in This represents the camera imaging calibration coefficient, measured in millimeters per pixel, and is determined based on the ratio of the actual physical size to the pixel size obtained from underwater checkerboard calibration experiments.

[0106] Using a pre-constructed allometric power function model relating carapace physical size to biomass, the estimated weight of each live crab was calculated based on physical size data. ,Right now ,in and These represent the conditional factors and growth index of allometric growth in crustaceans, respectively. They are derived by using historical records of the physical dimensions and precise weighing data of crab samples from this water area, and then employing a nonlinear regression analysis using the least squares method. , Around 3.0.

[0107] The average value and distribution statistics of the real-time stock, physical size data of all live crabs and estimated weight of each individual are calculated to generate a structured report containing population density, average size parameters and population weight distribution histogram, and this structured report is defined as growth monitoring and evaluation data.

[0108] Step S7 establishes a mapping relationship from image pixel space to physical biomass space, and utilizes the unique allometric growth law of crustaceans (power function relationship between length and weight) to achieve accurate digital assessment of the biomass and size distribution of the cultured population without contact.

[0109] Please see Figure 4 To verify the effectiveness of the fluid dynamics response analysis mechanism in distinguishing visually similar targets, this embodiment selects typical test samples for feature space mapping analysis. Figure 4This study demonstrates the clustering distribution of targets in different biological states within a two-dimensional dynamic feature space constructed from macroscopic drift cross-correlation and joint disturbance resistance damping characteristics. Live crabs, due to their active counter-current movement and appendage muscle stiffness, are mainly distributed in the upper left region of the feature space, exhibiting lower macroscopic drift cross-correlation and higher joint disturbance resistance damping characteristics. In contrast, dead crabs and crab molts are significantly dominated by the background microfluidic field, showing higher macroscopic drift cross-correlation. However, thanks to the microscopic physical differences in their physical mass distribution and joint connection tightness, they exhibit a distinguishable hierarchical distribution in the joint disturbance resistance damping characteristic dimension. This data analysis intuitively demonstrates that the method proposed in this invention can effectively overcome the limitations of traditional visual features in turbid dynamic water bodies, achieving high-precision dynamic decoupling and classification of three types of targets. Specifically:

[0110] By constructing a background microfluidic field vector to quantify the dynamic fluid excitation of the aquaculture environment, and simultaneously calculating the macroscopic drift cross-correlation between the displacement vector of the target centroid and the background flow field, the flow-dependent motion characteristics of the target are deconstructed (corresponding to...). Figure 4 The distribution along the mid-horizontal axis, combined with the joint disturbance resistance damping characteristics of the appendiceal joints under fluid impact, characterizes the differences in limb mechanical stiffness (corresponding to...). Figure 4 (Distributed along the central vertical axis), and finally, using a manifold coupled visual dynamics model, these features reflecting the physical motion mechanism are mapped in high dimension, thereby achieving accurate identification and effective clustering of live crabs, dead crabs, and crab molts under the condition of high visual similarity.

[0111] As can be seen from the above description, the method for monitoring and evaluating the growth status of river crabs provided in this embodiment has the following technical effects:

[0112] By accurately locating underwater targets and constructing background microfluidic field vectors, a target fluid dynamics response analysis mechanism based on microfluidic field excitation was established. Utilizing the skeletal key point features including the target centroid and appendage joints, the macroscopic drift cross-correlation characteristic representing the degree of drift with the current and the joint disturbance resistance damping characteristics reflecting the mechanical stiffness of the limbs were calculated. This effectively solved the technical challenge of distinguishing between live crabs, crab molts, and dead crabs in dynamic water bodies due to their highly similar visual appearance. Combining the high-dimensional nonlinear classification capability of the manifold-coupled visual dynamics model, the biological state of the target was accurately determined, and growth monitoring and evaluation data including population density and size parameters were generated based on allometric growth patterns. This enabled non-contact intelligent monitoring and scientific evaluation of the crab farming process.

[0113] Please see Figure 5The second objective of this embodiment is to provide a system for monitoring and evaluating the growth status of crabs, including an image acquisition and target localization module 100, a background microfluidic field construction module 200, a skeleton key point extraction module 300, a macroscopic drift cross-correlation analysis module 400, a joint disturbance resistance damping characteristic calculation module 500, a manifold coupled visual dynamics recognition module 600, and a growth monitoring and statistical evaluation module 700, wherein:

[0114] The image acquisition and target localization module 100 acquires a sequence of images of the aquaculture area and locates the target to be monitored in the sequence of images;

[0115] The background microflow field construction module 200 extracts the motion trajectory of suspended particles in the background region of the sequence image and constructs a temporal background microflow field vector;

[0116] The skeleton key point extraction module 300 extracts the skeleton key points of the target to be monitored, including the target centroid point and appendage joint points;

[0117] The macroscopic drift cross-correlation analysis module 400 performs fluid dynamic response analysis on the target under monitoring based on the background micro-flow field vector, and calculates the macroscopic drift cross-correlation between the displacement vector of the target centroid point and the background micro-flow field vector.

[0118] The joint disturbance resistance and damping characteristic calculation module 500 tracks the temporal micro-motion trajectory of the appendage joint relative to the trunk connection point, and calculates the joint disturbance resistance and damping characteristics of the appendage joint by combining the excitation effect of the background micro-flow field vector.

[0119] The manifold-coupled visual dynamics recognition module 600 inputs the macroscopic drift cross-correlation and joint disturbance rejection damping characteristics into the manifold-coupled visual dynamics model and outputs the biological state of the target to be monitored, including live crabs, crab molts, and dead crabs.

[0120] The growth monitoring and statistical evaluation module 700 statistically determines the target number and morphological parameters of live river crabs and generates growth monitoring and evaluation data.

[0121] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for monitoring and evaluating the growth status of river crabs, characterized in that, The methods and steps include the following: S1. Acquire a sequence of images of the aquaculture area and locate the target to be monitored in the sequence of images; S2. Extract the motion trajectory of suspended particles in the background region of the sequence image to construct a temporal background microfluidic field vector; S3. Extract the key points of the skeleton of the target to be monitored, including the target centroid and the joint points of the appendages; S4. Perform fluid dynamics response analysis on the target to be monitored based on the background micro-flow field vector, and calculate the macroscopic drift cross-correlation between the displacement vector of the target centroid point and the background micro-flow field vector; S5. Track the temporal micro-motion trajectory of the appendage joint relative to the trunk connection point, and calculate the joint disturbance resistance damping characteristics of the appendage joint in combination with the excitation effect of the background micro-flow field vector. S6. Input the macroscopic drift cross-correlation and the joint disturbance rejection damping characteristics into the manifold coupled visual dynamics model, and output the biological state of the target to be monitored, including live crabs, crab molts and dead crabs. S7. Statistically determine the target number and morphological parameters of live river crabs, and generate growth monitoring and evaluation data.

2. The method for monitoring and evaluating the growth status of river crabs according to claim 1, characterized in that, The process of locating the target to be monitored specifically includes: The underwater optical camera continuously collects primary analog video signals of the underwater environment using a fixed bracket installed at the bottom of the aquaculture area, and converts the primary analog video signals into digital sequence images using a video capture card; The digital sequence image is subjected to pixel-level enhancement processing to generate an enhanced digital sequence image; A convolutional neural network model containing a deep residual operation module and a multi-scale feature fusion structure is constructed. The enhanced digital sequence image is input into the convolutional neural network model, and the bounding box coordinates of the target to be monitored in the current frame image and the target confidence score are output. The Kalman filter multi-target tracking algorithm is used to perform cross-frame trajectory association and assign a unique identification code to the target to be monitored in the digital sequence image based on the intersection-union ratio of the bounding box coordinates between adjacent frames and the association cost matrix, thereby completing the continuous localization of the target to be monitored.

3. The method for monitoring and evaluating the growth status of river crabs according to claim 2, characterized in that, The process of constructing the background microfluidic field vector specifically includes: Based on the bounding box coordinates of each target to be monitored in the current frame, a target inversion mask is generated in the enhanced digital sequence image, and the image region in the enhanced digital sequence image excluding the area covered by the target inversion mask is defined as the background region of interest. The background region of interest is segmented using a Gaussian mixture background modeling algorithm to extract dynamically changing pixel groups, and the dynamically changing pixel groups whose pixel area is within a preset threshold are identified as floating particle feature points. The pyramid Lucas-Carnard sparse optical flow algorithm is used to calculate the pixel displacement vector of each suspended particle feature point between the current frame and the next frame. The effective pixel displacement vectors after removing outliers are calculated by vector weighted average to obtain the background average flow velocity vector and background average flow direction vector of the current frame. The background average velocity vector and background average flow direction vector calculated for each frame in the continuous time series are combined in time stamp order to construct a temporal background microflow field vector.

4. The method for monitoring and evaluating the growth status of river crabs according to claim 2, characterized in that, The extraction process of the key points of the skeleton specifically includes: Based on the bounding box coordinates of each target to be monitored in the current frame, the region of interest image of the target to be monitored is extracted from the enhanced digital sequence image; The image of the region of interest of the target to be monitored is input into a cascaded pyramid pose estimation neural network model that has been pre-trained and converged based on a dataset of key points labeled on the skeleton of a crab. The cascaded pyramid-shaped attitude estimation neural network model is used to output a two-dimensional Gaussian response heatmap corresponding to multiple channels of the anatomical structure of the target to be monitored; For each channel, the maximum coordinate search is performed on the two-dimensional Gaussian response heatmap. The pixel position with the highest response value is extracted as the key point coordinate. The pixel position corresponding to the geometric center of the dorsal side of the crab's cephalothorax is marked as the target centroid point. The pixel position corresponding to the joint connection of the crab's two walking legs and chelipeds is marked as the appendage joint point, thus completing the extraction of the skeletal key points of the target to be monitored.

5. The method for monitoring and evaluating the growth status of river crabs according to claim 3, characterized in that, The calculation process of the macroscopic drift cross-correlation specifically includes: Based on the pixel coordinates of the target centroid point in the continuous time series, the displacement vector of the target centroid point in the current frame relative to the previous frame is calculated and defined as the target centroid point displacement vector. Call the background average velocity vector and background average flow direction vector corresponding to the current frame from the background microflow field vector; Calculate the cosine of the angle between the target centroid point displacement vector and the background average flow direction vector, and use it as the flow direction consistency coefficient; The ratio of the magnitude of the displacement vector of the target centroid point to the magnitude of the average velocity vector of the background is calculated and used as the velocity response ratio coefficient. A linear weighted summation algorithm is used to fuse the flow direction consistency coefficient and the flow velocity response ratio coefficient to calculate the macroscopic drift cross-correlation that characterizes the degree of drift of the monitored target with the flow.

6. The method for monitoring and evaluating the growth status of river crabs according to claim 4, characterized in that, The calculation process for the joint's disturbance rejection damping characteristics specifically includes: Define the target centroid point of the current frame as the origin of the coordinate system, and calculate the relative coordinate vector of each appendage joint point relative to the target centroid point in the current frame; The relative coordinate vectors of all frames in the continuous time series are arranged in the order of timestamps to construct the temporal micro-motion trajectory of the appendage joint relative to the trunk connection point; The standard deviation of the time-domain micro-motion trajectory within a preset time window is calculated and defined as the limb micro-motion response intensity; The background average velocity vector corresponding to the current frame in the background microflow field vector is acquired synchronously, and its modulus is calculated as the fluid excitation intensity. The ratio of the fluid excitation intensity to the micromotion response intensity of the appendage is calculated and determined as the joint disturbance resistance damping characteristic of the appendage joint.

7. The method for monitoring and evaluating the growth status of river crabs according to claim 6, characterized in that, The process of determining the biological state specifically includes: The macroscopic drift cross-correlation and the joint disturbance rejection damping characteristics are combined to construct a two-dimensional state feature vector, and then normalized. The processed two-dimensional state feature vector is input into a convergent manifold-coupled visual dynamics model pre-trained based on sample data with biological tags. The two-dimensional state feature vector is mapped to a high-dimensional feature space using the radial basis kernel function integrated within the manifold coupled visual dynamics model, and the geometric margin between the mapped feature vector and the preset optimal classification hyperplane is calculated. The category is determined based on the value of the geometric interval and its location within the decision region, thus identifying the biological state of the target to be monitored.

8. The method for monitoring and evaluating the growth status of river crabs according to claim 1, characterized in that, The process of generating the growth monitoring and evaluation data specifically includes: Targets that are determined to be live river crabs are selected, and the real-time inventory of live river crabs is counted. For each live crab, the bounding box coordinates of its current frame are called to calculate the shell morphology features, and the shell morphology features are converted into physical size data by combining the preset camera imaging calibration coefficients. Using a pre-constructed allometric power function model relating carapace physical size and biomass, the estimated weight of each live crab is calculated based on the physical size data. The average value and distribution statistics of the real-time stock, physical size data of all live crabs and estimated weight of each individual are calculated to generate a structured report containing population density, average size parameters and a histogram of population weight distribution, which is defined as growth monitoring and evaluation data.

9. The method for monitoring and evaluating the growth status of river crabs according to claim 2, characterized in that, The pixel-level enhancement process employs a dark channel prior image dehazing algorithm combined with a contrast-limited adaptive histogram equalization algorithm.

10. A system using the method for monitoring and evaluating the growth status of crabs according to any one of claims 1-9, characterized in that, The module includes an image acquisition and target localization module (100), a background microfluidic field construction module (200), a skeleton key point extraction module (300), a macroscopic drift cross-correlation analysis module (400), a joint disturbance resistance damping characteristic calculation module (500), a manifold coupled visual dynamics recognition module (600), and a growth monitoring and statistical evaluation module (700), wherein: The image acquisition and target positioning module (100) acquires a sequence of images of the aquaculture area and locates the target to be monitored in the sequence of images; The background microflow field construction module (200) extracts the motion trajectory of suspended particles in the background region of the sequence image and constructs a temporal background microflow field vector; The skeleton key point extraction module (300) extracts the skeleton key points of the target to be monitored, including the target centroid point and appendage joint points; The macroscopic drift cross-correlation analysis module (400) performs hydrodynamic response analysis on the target to be monitored based on the background micro-flow field vector, and calculates the macroscopic drift cross-correlation between the displacement vector of the target centroid point and the background micro-flow field vector. The joint disturbance resistance damping characteristic calculation module (500) tracks the temporal micro-motion trajectory of the appendage joint relative to the trunk connection point, and calculates the joint disturbance resistance damping characteristics of the appendage joint in combination with the excitation effect of the background micro-flow field vector. The manifold coupled visual dynamics recognition module (600) inputs the macroscopic drift cross-correlation and the joint disturbance rejection damping characteristics into the manifold coupled visual dynamics model and outputs the biological state of the target to be monitored, including live crabs, crab molts and dead crabs. The growth monitoring and statistical evaluation module (700) statistically determines the target number and morphological parameters of live river crabs and generates growth monitoring and evaluation data.

Images