Net cage fish grouping and directional feeding method and system based on stereoscopic vision

Abstract

The invention discloses a net cage fish grouping and directional feeding method and system based on stereoscopic vision. The method comprises the following steps: acquiring video data through a multi-view underwater camera system, generating three-dimensional point cloud data of a fish body based on a binocular vision principle, and extracting growth and behavior characteristics; dividing a target group by using a clustering algorithm, and generating a differentiated feeding strategy in combination with the environmental parameters; and driving the multi-outlet directional bait casting machine, and selecting a direct falling mode or an air conveying mode according to the water layer distribution identifier to deliver the feed to the target area. Precise sensing and layered feeding of fish school specifications are achieved, the problem of uneven ingestion caused by specification differentiation is solved, the feed utilization rate is increased, residual feed pollution is reduced, and tidy growth of fishes is promoted.

Description

Technical Field

[0001] This invention belongs to the field of intelligent fisheries and aquaculture automation technology, and relates to a method and system for grouping and directional feeding of cage-cultured fish based on stereoscopic vision. Background Technology

[0002] With the large-scale development of aquaculture, deep-sea cage culture has become an important model for increasing aquatic product yields. In cage culture, feeding management is a crucial factor determining farming costs, fish growth rates, and aquatic environmental safety. Traditional feeding methods often rely on manual experience or simple timed and quantitative mechanical distribution, typically employing a strategy of uniform distribution throughout the pond. However, due to natural genetic differences, varying feeding abilities, and environmental adaptability among individual fish, long-term polyculture can lead to significant size differentiation within the fish population.

[0003] Existing conventional feeding techniques have the following main drawbacks: First, they lack precise means to perceive the real-time distribution and individual size of fish schools. Traditional methods cannot identify the vertical stratification of fish schools in the water (e.g., small fry tend to gather in the middle and upper layers, while large adults tend to inhabit the bottom), leading to a one-size-fits-all feeding strategy. If the feeding amount is too small, the dominant large-sized groups have a strong ability to compete for food, while the smaller groups are disadvantaged in competition and their growth is hindered, further widening the size difference within the group and reducing the overall quality of the fish. If the feeding amount is too large to meet the needs of the smaller groups, it will result in a large amount of feed waste. Uneaten feed sinks to the bottom and rots, which not only increases breeding costs but also deteriorates water quality and induces fish diseases.

[0004] Secondly, existing automated feeding equipment has limited functionality, mostly possessing only simple flow control or horizontal rotation functions, lacking intelligent switching mechanisms for differentiated feeding modes at different water depths. Although some studies have attempted to introduce underwater camera technology, these are mostly limited to two-dimensional planar image analysis, making it difficult to accurately obtain the three-dimensional spatial position, true body size, and complex group behavior characteristics of fish. This results in low accuracy in grouping judgment and an inability to support refined targeted feeding decisions.

[0005] Therefore, it is necessary to develop a method and system that can accurately acquire the three-dimensional characteristics of fish schools using stereoscopic vision technology, automatically classify target groups of different sizes, and dynamically adjust feeding patterns and strategies according to the water layer where the group is located, so as to achieve precise, intelligent and ecological management of cage aquaculture. Summary of the Invention

[0006] To address the problems existing in the background technology, this invention proposes a method and system for grouping and directional feeding of fish in net cages based on stereoscopic vision.

[0007] To achieve the above objectives, the technical solution adopted by this invention is as follows: a method for grouping and directional feeding of fish in net cages based on stereoscopic vision, comprising the following steps: The system collects video data of fish schools using a multi-view underwater camera system, and processes the video data based on the principle of binocular vision to generate three-dimensional point cloud data of the fish bodies. The body length and weight data of each individual fish are extracted from the three-dimensional point cloud data of the fish body as growth characteristics. Combined with behavioral characteristic data, a clustering algorithm is used to divide the fish group into at least two target groups with different growth characteristics. Based on the growth characteristics of each target population and current environmental parameters, generate differentiated feeding strategies that include feed type, feeding amount and feeding area; The multi-outlet directional feeder is controlled to accurately deliver feed to the activity areas of each target group according to the differentiated feeding strategy.

[0008] Specifically, the combination of behavioral feature data includes: A deep learning model is used to identify the biological characteristics of fish, and an inter-frame matching algorithm is used to continuously track each individual fish in consecutive video frames to obtain swimming speed and feeding frequency as behavioral characteristic data.

[0009] Specifically, the method of using clustering algorithms to divide the fish population into at least two target groups with different growth characteristics includes any of the following: Method 1: Calculate the average body length and classify fish individuals whose actual body length is greater than, within, or less than the preset proportion of the average body length into fast-growing, normal-growing, or slow-growing groups. Method 2: Using body length as the core feature, construct a multi-dimensional feature space by combining water temperature, dissolved oxygen, salinity, pH value, and flow rate, and use the K-means clustering algorithm to dynamically divide the fish population into a preset number of groups or groups whose number is automatically determined based on clustering effectiveness indicators.

[0010] Specifically, the generation of differentiated feeding strategies includes: The nutritional requirement model, trained based on fish species data, grouping characteristics, and seasonal rhythms, is used to calculate the target feeding amount for each target group. The fitness analysis model is used to quantitatively analyze the matching relationship between environmental parameters and feed nutrients, output feed fitness scores and match feeds with corresponding nutrient ratios. After performing a logical consistency check on the feeding strategy data, a standardized feeding instruction containing water layer distribution identifiers is generated.

[0011] Specifically, the control of the multi-outlet directional feeder, according to the differentiated feeding strategy, precisely delivers feed to the activity areas of each target group, including: Parse the water layer distribution identifier in the standardized feeding instructions; If the identifier indicates that the target group is located at the bottom of the cage, then the feeder is controlled to execute a direct-fall feeding mode; If the identifier indicates that the target group is distributed in the upper layer of the cage, then the feeder is controlled to execute the wind-driven feeding mode.

[0012] Specifically, the direct-feed mode includes adjusting the flow regulating valve of the vertical feeding channel to control the feeding rate; The air-feeding mode includes adjusting the air pressure of the high-pressure blower and the airflow regulating valve to control the airflow intensity.

[0013] Specifically, the stereo vision-based method for grouping and directing fish in net cages also includes a feedback adjustment step: Periodically collect growth data and feed residue after feeding, and compare the actual and expected growth rates; If the deviation is lower than the preset threshold, the feeding parameters are adjusted and the nutritional requirement model is updated.

[0014] This technical solution also provides a stereo vision-based method for grouping and targeted feeding of fish in net cages, including: The stereo vision recognition module is configured to acquire video data through a multi-view underwater camera system and generate three-dimensional point cloud data of the fish body based on the principle of binocular vision, from which growth characteristics and behavioral characteristics data are extracted; The intelligent decision-making module is configured to divide the target population based on the feature data using a clustering algorithm, and generate differentiated feeding strategies based on growth characteristics and environmental parameters. The group feeding execution module includes a multi-outlet directional feeder, configured to deliver feed to the target area according to the differentiated feeding strategy; The central processing unit is used to schedule data interaction and instruction execution among various modules.

[0015] Specifically, the intelligent decision-making module includes a nutrient requirement analysis unit, a feed adaptation unit, and an instruction generation unit, which are respectively configured to calculate the target feeding amount, match the feed nutrient ratio, and generate standardized feeding instructions containing water layer distribution indicators.

[0016] Specifically, the feature is that the group feeding execution module further includes a group feeding controller, configured to parse the water layer distribution marker, generate a first control signal to drive the feeder to execute a direct-fall feeding mode when the marker indicates the bottom layer, and generate a second control signal to drive the feeder to execute a wind-driven feeding mode when the marker indicates the middle and upper layers.

[0017] Compared with existing technologies, this invention has the following advantages: First, it achieves precise three-dimensional perception and dynamic grouping of fish schools. Through a multi-view underwater camera system and binocular vision principles, this system can generate high-precision three-dimensional point cloud data of fish bodies, overcoming the shortcomings of traditional two-dimensional vision due to light refraction and occlusion, and accurately extracting the true growth and behavioral characteristics of fish. Based on this data, clustering algorithms are used to automatically divide target groups, enabling real-time monitoring of fish size differentiation and water layer distribution within the cages, providing reliable data support for differentiated feeding.

[0018] Secondly, it improves feed utilization and reduces aquaculture costs. The system generates differentiated feeding strategies based on the growth characteristics and environmental parameters of each target group, and achieves precise delivery through a multi-outlet directional feeder. A drop-feed method is used for large bottom-dwelling fish, while a wind-driven diffusion method is used for small fish in the middle and upper layers, avoiding the problem of dominant groups over-feeding while weaker groups under-feed due to a one-size-fits-all approach. This not only improves feed conversion rate and promotes uniform fish size, but also reduces water pollution from uneaten feed settling at the bottom, lowering the risk of disease.

[0019] Third, it enhances the system's intelligent decision-making and adaptive capabilities. Through a closed-loop feedback mechanism, the system can automatically adjust the nutrient requirement model parameters based on the actual growth rate, ensuring continuously optimized feeding results. Attached Figure Description

[0020] Figure 1 This is a flowchart of the method for grouping and targeted feeding of fish in net cages based on stereoscopic vision according to the present invention; Figure 2 This is a block diagram of the cage fish grouping and targeted feeding system based on stereoscopic vision of the present invention; Figure 3 This is a power function fitting curve of the relationship between body length and weight according to the present invention. Detailed Implementation

[0021] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] like Figures 1-3 As shown, the technical solution adopted in this invention is as follows: a method for grouping and targeted feeding of fish in net cages based on stereoscopic vision, comprising the following steps: S1: Collect video data of fish schools through a multi-view underwater camera system, and process the video data based on the principle of binocular vision to generate three-dimensional point cloud data of the fish.

[0023] The multi-view underwater camera system refers to an image acquisition device consisting of three high-definition waterproof cameras. The three high-definition waterproof cameras are installed on the inner wall of the net cage by a bracket and are arranged in a triangular pattern. The lens surface is coated with an anti-fog and anti-reflection coating, and the protection level reaches the IP68 standard to adapt to the 50-meter water depth environment, ensuring coverage of more than 95% of the activity area inside the net cage. Its function is to capture the original video stream of fish activity in real time.

[0024] Fish school video data refers to a sequence of digital images containing information on the movement trajectory and morphology of fish schools, which are continuously captured by a multi-view underwater camera system at a preset cycle.

[0025] The principle of binocular vision refers to a geometric measurement method that uses two or more cameras to simultaneously capture the same scene from different angles, and then calculates the parallax of the same object in images from different viewpoints to recover the three-dimensional spatial coordinates of the object.

[0026] Fish body 3D point cloud data refers to a data model generated by matching feature points and triangulation calculations on stereo image pairs in fish video data based on the principle of binocular vision. It consists of a large set of points with three-dimensional spatial coordinates, where the points represent the horizontal, vertical and depth positions of the fish body surface points in the net cage coordinate system, with the unit being millimeters (mm). This data model is used to accurately reconstruct the three-dimensional geometric shape of the fish body.

[0027] The multi-view underwater camera system is activated and collects video data of the fish school. After receiving the video data, the central processing unit calls the 3D reconstruction algorithm in the stereo vision recognition module. Based on the principle of binocular vision, it performs stereo matching calculation on the video frames to calculate the depth information of the fish surface and finally generates 3D point cloud data of the fish.

[0028] S2: Extract the body length and weight data of each individual fish from the three-dimensional point cloud data of the fish body as growth characteristics, and combine them with behavioral characteristic data to divide the fish group into at least two target groups with different growth characteristics using a clustering algorithm.

[0029] Fish body 3D point cloud data refers to a data model consisting of a large set of points with 3D spatial coordinates generated in the preceding steps.

[0030] Body length data refers to the physical quantity obtained by calculating the maximum Euclidean distance of the point cloud model of the i-th individual fish in the principal axis direction of the three-dimensional point cloud data of the fish body. The unit is millimeters (mm). The mathematical expression is , where and are located at both ends of the principal axis.

[0031] Weight data refers to the physical quantity of mass obtained by calculating the volume of the fish based on the three-dimensional point cloud data of the fish body for the i-th individual, and combining it with the fish density constant (typical value is taken, based on the engineering experience value of the average density of general freshwater fish), with the unit being grams (g), and the mathematical expression being.

[0032] Growth characteristics refer to a two-dimensional vector composed of body length and weight data, which is used to characterize the growth and development status of the i-th individual fish.

[0033] Behavioral feature data refers to the data set that characterizes the movement and feeding habits of an individual fish. Clustering algorithms are unsupervised machine learning methods whose input is a set of feature vectors containing growth and behavioral feature data, and whose output is a set of labels dividing the input data into clusters.

[0034] The target group refers to a set of fish individuals with similar growth characteristics, identified through clustering algorithms.

[0035] The specific implementation logic of this step is as follows: The central processing unit reads the three-dimensional point cloud data of the fish body, calculates the body length and weight data of each individual to construct growth characteristics, and then concatenates the growth characteristics with the acquired behavioral characteristic data to form a comprehensive feature vector, which is input to the clustering algorithm module. If the clustering algorithm is completed, it outputs at least two target groups.

[0036] Specifically, the combination of behavioral feature data includes: using a deep learning model to identify the biological characteristics of the fish, and continuously tracking each individual fish in consecutive video frames using an inter-frame matching algorithm to obtain swimming speed and feeding frequency as the behavioral feature data.

[0037] A deep learning model is a software module based on a convolutional neural network (CNN) or recurrent neural network (RNN) architecture. Its input is a single frame or multiple frames of fish video images, and its output is the biometric identifier of individual fish and the coordinates of key points. The biometric identifier is used to uniquely distinguish different individual fish.

[0038] Inter-frame matching algorithm refers to a computational logic used to associate the same target in a time series. Its input is the sequence of key point coordinates output by a deep learning model in consecutive video frames, and the output is the trajectory chain of the same fish individual at different times.

[0039] Swimming speed refers to the displacement of the i-th individual fish per unit time, measured in meters per second (m / s). The calculation formula is: where represents the spatial coordinates of the i-th individual fish at time t, and is the time interval between video frames. Feeding frequency refers to the number of times the i-th individual fish performs feeding actions per unit time, measured in times per minute (times / min). It is obtained by statistically analyzing the frequency of mouth opening and closing actions or specific feeding postures identified by a deep learning model.

[0040] The specific implementation logic is as follows: The system calls a deep learning model to process the current video frame and outputs the biological feature identifiers of each individual fish; then, it executes an inter-frame matching algorithm. If the biological feature identifier in the current frame matches the biological feature identifier in the previous frame, the trajectory of the individual is updated and the swimming speed is calculated; at the same time, if the deep learning model detects feeding actions, it accumulates the count to calculate the feeding frequency; finally, the swimming speed and feeding frequency are combined to form behavioral feature data.

[0041] Specifically, the method of using clustering algorithms to divide the fish population into at least two target groups with different growth characteristics includes any of the following: Method 1: Calculate the average body length and divide fish individuals whose actual body length is greater than, within, or less than the preset proportion of the average body length into fast, normal, or slow growth groups.

[0042] The average body length value refers to the arithmetic mean of the body length data of all identified fish individuals in the current net cage. The mathematical expression is: , where is the total number of fish individuals, and is the body length data of the i-th individual.

[0043] The preset ratio range refers to a pre-set positive real number percentage threshold used to define the fluctuation range of normal growth. The typical value is 10% (i.e., based on the typical engineering experience value of the growth difference of fish in the same batch in aquaculture).

[0044] The rapid growth group refers to a collection of fish individuals that meet the criteria; the normal growth group refers to a collection of fish individuals that meet the criteria.

[0045] The slow-growth group refers to the set of fish individuals that meet the criteria; the central processing unit first calculates the average body length. Then, it iterates through all fish individuals, and for the i-th fish individual, if the body length is greater than a certain value, then the i-th fish individual is classified into the fast-growth group.

[0046] If the body length data is less than 1, then the i-th fish individual will be classified into the slow growth group.

[0047] If the body length data is within the range, the first fish individual will be classified into the normal growth group.

[0048] Method 2: Using body length as the core feature, a multi-dimensional feature space is constructed by combining water temperature, dissolved oxygen, salinity, pH value and flow rate. The K-means clustering algorithm is used to dynamically divide the fish population into groups of a preset number or groups whose number is automatically determined based on the clustering effectiveness index.

[0049] Water temperature refers to the real-time temperature of the water in net cage aquaculture, measured in degrees Celsius (C), and is collected by temperature sensors deployed underwater.

[0050] Dissolved oxygen refers to the concentration of oxygen dissolved in water, measured in milligrams per liter (mg / L), and is collected by a dissolved oxygen sensor.

[0051] Salinity refers to the total amount of dissolved salts in a body of water, expressed in parts per thousand (ppm), and is measured using a salinity meter.

[0052] pH value is a dimensionless quantity referring to the acidity or alkalinity of water, and it is collected by a pH sensor.

[0053] Flow velocity refers to the speed at which water flows, measured in meters per second (m / s), and is collected by a flow meter.

[0054] The multidimensional feature space refers to a vector space composed of six dimensions: body length data, water temperature, dissolved oxygen, salinity, pH value, and flow velocity. The feature vector of the i-th fish individual in this space is .

[0055] K-means clustering is an iterative clustering analysis algorithm. Its inputs are a set of feature vectors and the number of clusters, and its outputs are a set of cluster labels and the center coordinates of each cluster.

[0056] The preset number refers to the number of clusters specified by the user in advance, which is a positive integer; the clustering effectiveness index is a quantitative indicator used to evaluate the quality of clustering results, typically the silhouette coefficient, which ranges from [-1, 1], and the closer the value is to 1, the better the clustering effect.

[0057] Automatically determining the number of clusters refers to iterating through different candidate cluster numbers (for example), running the K-means clustering algorithm on each one and calculating the corresponding clustering effectiveness index, and selecting the value that maximizes the clustering effectiveness index as the final number of clusters.

[0058] The system collects environmental parameters to construct feature vectors in a multidimensional feature space. If the user specifies a preset number, the number of clusters is set; otherwise, the system performs a loop calculation, initializes the maximum index value, and for each candidate value, runs the K-means clustering algorithm to generate a temporary label set and calculates the current clustering effectiveness index. If the index is positive, the number of clusters is updated and set. After the loop ends, the K-means clustering algorithm is run again using the final determined number of clusters to generate the final cluster label set, thus dividing the fish population into the specified number of clusters.

[0059] S3: Generate differentiated feeding strategies that include feed type, feeding amount and feeding area based on the growth characteristics of each target group and current environmental parameters.

[0060] The target group refers to the set of fish individuals with similar growth characteristics identified by clustering algorithms in the previous steps, where is the number of clusters. Growth characteristics refer to the average growth feature vector obtained statistically for the i-th target group, including average body length, average weight, and average behavioral characteristics, used to characterize the overall developmental state of the group. Current environmental parameters refer to the set of water temperature, dissolved oxygen, salinity, pH value, and flow velocity data collected in real time by sensors, i.e., where the definitions and units of each parameter are as described above.

[0061] Differentiated feeding strategy refers to the set of control instructions generated for the i-th target group. Its structure is defined as a triple, where: represents the feed type, which refers to the nutrient ratio code of the feed; represents the feeding amount, in grams (g), which refers to the planned feed mass; and represents the feeding area, which refers to the three-dimensional spatial coordinate range within the cage.

[0062] The central processing unit reads the growth characteristics of all target populations and the latest current environmental parameters, calls the strategy generation module, and calculates the corresponding and for each population to generate a differentiated feeding strategy for each population.

[0063] Specifically, the generation of differentiated feeding strategies includes: S3.1: Call the nutritional requirement model trained based on fish species basic data, group characteristics and seasonal rhythms to calculate the target feeding amount for each target group.

[0064] Basic data on fish species refers to a set of biological constants about farmed fish species that are stored in advance in a database, including basal metabolic rate, optimal growth temperature range and nutrient conversion coefficient. The data sources are industry-recognized aquaculture manuals or experimental measurements.

[0065] Cluster characteristics refer to the statistical characteristic vector of the i-th target group, which includes the average body length, average weight, and activity intensity index of all individuals in the group. Seasonal rhythm refers to the correction coefficient reflecting the influence of the current date and season on fish feeding, with a value range of [0,1]. Typical values ​​are: 0.8 for spring, 1.0 for summer, 0.9 for autumn, and 0.5 for winter, based on typical engineering experience values ​​of the feeding habits of common farmed fish in temperate regions.

[0066] A nutritional requirement model is a software module that has been trained offline. It is implemented by a regression neural network or a multiple linear regression equation. The inputs are basic data of fish species, grouping characteristics and seasonal rhythms, and the output is the target feeding amount.

[0067] The target feeding amount refers to the feed mass required by the target population per unit time period, expressed in grams (g).

[0068] The system obtains the seasonal factors of the current date to determine the seasonal rhythm, extracts the grouping characteristics of each population, loads the preset basic data of fish species, and inputs the three into the nutritional requirement model; if the nutritional requirement model is completed, it outputs the corresponding target feeding amount and assigns this value to the feeding amount in the differentiated feeding strategy.

[0069] S3.2: Use the fit analysis model to quantitatively analyze the matching relationship between environmental parameters and feed nutrients, output the feed fit score and match the feed with the corresponding nutrient ratio.

[0070] Feed nutrient composition refers to the vector of chemical components of each feed in the candidate feed library, including protein content, fat content, carbohydrate content, and vitamin content, expressed as a percentage (%).

[0071] The fit analysis model is a computational module used to assess the degree of matching between environmental conditions and feed characteristics. Its inputs are current environmental parameters and the feed nutrient composition of candidate feeds, and its output is the feed fit score.

[0072] Feed fit score is a dimensionless scalar with a value range of [0,1], where 1 represents a perfect match and 0 represents a complete mismatch. The typical calculation formula uses weighted cosine similarity: where and are the normalized environmental parameter vector and nutrient component vector, respectively. Feed type refers to the feed code selected from the candidate feed library that has specific feed nutrient components.

[0073] The system iterates through each feed in the candidate feed library, extracts its nutritional components, and inputs them along with the current environmental parameters into the fitness analysis model. For the i-th candidate feed, if the calculated feed fitness score is greater than the scores of all other candidate feeds, i.e., the code of the i-th feed is selected as the feed type and assigned to the corresponding field in the differentiated feeding strategy.

[0074] S3.3: After performing logical consistency verification on the feeding strategy data, generate a standardized feeding instruction containing water layer distribution identifiers.

[0075] Feeding strategy data refers to a collection of differentiated feeding strategies for all target groups.

[0076] Logical consistency verification refers to a set of Boolean rules used to verify the validity of feeding strategy data, including: whether the total feeding amount exceeds the cage's carrying capacity threshold, whether the feeding area is within the cage's physical boundaries, and whether the feed type exists in the available inventory. Standardized feeding instructions refer to the final control data packet conforming to the communication protocol format, whose structure includes header information, strategy payload, and checksum.

[0077] The water layer distribution identifier is a key field in the standardized feeding instructions, used to indicate the operating mode of the feeder. It is defined as an integer enumeration: if , it indicates direct drop feeding (for bottom-dwelling fish); if , it indicates wind-driven feeding (for mid-to-upper-level fish). This mapping relationship is set according to the publicly available technical documents of the feeder manufacturer.

[0078] The central processing unit performs a logical consistency check on the feeding strategy data; if all check rules pass (i.e., the total feeding amount threshold, the area coordinate cage boundary, and the feed type inventory list), the system determines the water layer distribution identifier based on the depth information of the feeding area in the differentiated feeding strategy.

[0079] If the average depth is greater than the total depth of the cage, then set it; otherwise, set it. Subsequently, the system encapsulates the feed type, feeding amount, and determined water layer distribution identifier to generate standardized feeding instructions. If any verification rule fails, the system reports an error and stops instruction generation, maintaining the feeding strategy of the previous cycle or entering a safe shutdown mode.

[0080] S4: Control the multi-outlet directional feeder to accurately deliver feed to the activity area of ​​each target group according to the differentiated feeding strategy.

[0081] A multi-outlet directional feeder is a hardware device with multiple independent discharge ports, each with its own controllable feeding method. Its structure includes a central feed hopper, a vertical feeding channel, a high-pressure blower system, a flow control valve group, and an airflow control valve group, used to execute physical feeding actions. A differentiated feeding strategy refers to the set of control instructions generated in the preceding steps for a specific target group, including feed type, feeding amount, and feeding area.

[0082] The activity area refers to the three-dimensional spatial range in which the target group gathers at the current moment, which is determined in real time by the preceding visual tracking system.

[0083] The central processing unit reads the differentiated feeding strategies for all target groups and the activity areas of each group, and generates a control signal sequence to send to the multi-outlet directional feeder. The multi-outlet directional feeder receives the control signal, drives the corresponding discharge port to move, and projects the specified weight of feed to the designated activity area, thereby achieving precise delivery.

[0084] Specifically, the control of the multi-outlet directional feeder, according to the differentiated feeding strategy, precisely delivers feed to the activity areas of each target group, including: The water layer distribution identifier in the standardized feeding instructions is analyzed.

[0085] If the identifier indicates that the target group is located at the bottom of the cage, then the feeder is controlled to execute a direct-drop feeding mode.

[0086] If the identifier indicates that the target group is distributed in the upper layer of the cage, then the feeder is controlled to execute the wind-driven feeding mode.

[0087] The standardized feeding instruction refers to the final control data packet generated in the preceding steps that conforms to the communication protocol format, which includes a water layer distribution identifier field. The water layer distribution identifier is an integer enumeration variable, defined as follows: if , it means that the target group is distributed at the bottom of the cage; if , it means that the target group is distributed in the upper layer of the cage.

[0088] The central processing unit receives standardized feeding instructions, extracts the water layer distribution identifier values ​​through the data parsing module, and temporarily stores the values ​​in a register for subsequent logical judgment.

[0089] The direct-fall feeding mode refers to a feeding mechanism that uses gravity to allow feed to fall freely into the water along a vertical channel, which is suitable for bottom-dwelling fish.

[0090] The central processing unit determines the value of the extracted water layer distribution indicator; if the water layer distribution indicator is equal to 1, the central processing unit sends a mode switching command to the multi-outlet directional feeder, setting the current working mode to the direct drop feeding mode; subsequently, the multi-outlet directional feeder starts the vertical feeding channel, shuts down the high-pressure blower system, and prepares to perform gravity feeding.

[0091] The wind-driven feeding mode refers to a feeding mechanism that uses high-pressure airflow to blow and diffuse feed to the upper and middle layers of the water, which is suitable for feeding fish in the middle and upper layers.

[0092] The central processing unit determines the value of the extracted water layer distribution indicator; if the water layer distribution indicator is equal to 2, the central processing unit sends a mode switching command to the multi-outlet directional feeder, setting the current working mode to the air-feeding mode; subsequently, the multi-outlet directional feeder starts the high-pressure fan system, opens the airflow channel, and prepares to perform airflow feeding.

[0093] Specifically, the direct-feed mode includes adjusting the flow regulating valve of the vertical feeding channel to control the feeding rate.

[0094] The air-feeding mode includes adjusting the air pressure of the high-pressure blower and the airflow regulating valve to control the airflow intensity.

[0095] A vertical feed channel refers to the physical pipe in a multi-outlet directional feeder used to guide feed to fall vertically. A flow control valve is an electric or pneumatic valve installed on the vertical feed channel; its opening range is [value missing], used to regulate the amount of feed passing through per unit time.

[0096] Feeding rate refers to the mass of feed passing through the vertical feeding channel per unit time, measured in grams per second (g / s). It is positively correlated with the opening degree, with a typical mapping relationship as follows: where is the valve flow coefficient (typical values ​​are determined based on the valve manufacturer's technical documentation).

[0097] When the system is in direct-fall feeding mode, the central processing unit calculates the target feeding rate based on the feeding amount and preset feeding duration in the differentiated feeding strategy. Subsequently, the central processing unit calculates the required valve opening and controls the flow regulating valve to rotate to that opening, thereby precisely controlling the feed falling speed.

[0098] A high-pressure blower is a power component in a multi-outlet directional feeder used to generate high-speed airflow. Its output parameter is air pressure, measured in Pascals (Pa). An airflow regulating valve is a regulating device installed on the outlet duct, with an opening range of [0%, 100%], used to adjust the cross-sectional area of ​​the airflow.

[0099] Airflow intensity refers to the kinetic energy flux of gas passing through the air outlet per unit time, measured in watts per square meter. It is determined by both air pressure and the opening of the airflow regulating valve, with a typical relationship being as follows.

[0100] When the system is in pneumatic feeding mode, the central processing unit calculates the target airflow intensity based on the distance to the feeding area and the amount of feed. Subsequently, the central processing unit adjusts the speed of the high-pressure blower to set the target air pressure and controls the airflow regulating valve to open to the target opening degree, so that the synthesized airflow intensity can accurately deliver the feed to the target water layer.

[0101] Specifically, the stereo vision-based method for grouping and directing fish in net cages also includes a feedback adjustment step: Periodically collect growth data and feed residue after feeding, and compare the actual growth rate with the expected growth rate.

[0102] If the deviation is lower than the preset threshold, the feeding parameters are adjusted and the nutritional requirement model is updated.

[0103] The feedback adjustment step refers to the closed-loop optimization process executed after the main process of this method is completed; growth data refers to the statistical values ​​of fish body length and weight collected periodically (typically 7 days, based on the routine monitoring frequency of aquaculture) through the preceding stereo vision system. Feed residue refers to the mass of uneaten feed detected by underwater camera image analysis or bottom sensors after a single feeding cycle, measured in grams (g).

[0104] The actual growth rate refers to the increase in average body weight of fish per unit time, and is calculated using the formula: where and are the average body weight of the group at times and , respectively; the expected growth rate refers to the theoretical growth rate predicted by the preceding nutrient requirement model based on the current feeding strategy.

[0105] The system initiates feedback adjustment steps at preset intervals, collects the latest growth data, and calculates the remaining feed amount. The system uses the collected data to calculate the actual growth rate and retrieves the expected growth rate for the same time period from the database. Subsequently, the system calculates the difference between the two.

[0106] Deviation refers to the absolute value of the difference between the actual growth rate and the expected growth rate, expressed mathematically as follows. The preset threshold is the error limit for determining whether the model needs updating; a typical value is 50 mg per day, based on engineering experience regarding acceptable growth prediction errors in high-density farming environments. Feeding parameters refer to the set of adjustable variables affecting feeding effectiveness, including the basal metabolic rate coefficient, temperature correction factor, and feeding activity weight.

[0107] The central processing unit determines the relationship between the calculated deviation and the preset threshold. If the deviation is less than the preset threshold, it indicates that the current model prediction is accurate, and the system maintains the existing parameters unchanged. Otherwise (i.e., the deviation is greater than or equal to the preset threshold), the system performs parameter adjustment operations: using the gradient descent method or the least squares method, the correlation coefficient in the feeding parameters is corrected in reverse according to the actual growth rate and the amount of feed remaining.

[0108] Subsequently, the system writes the corrected feeding parameters into the parameter storage area of ​​the nutrient requirement model, completing the model update so that the updated nutrient requirement model can be used to generate a more accurate feeding strategy in the next feeding cycle.

[0109] This technical solution also provides a stereo vision-based method for grouping and targeted feeding of fish in net cages, including: The stereo vision recognition module is configured to acquire video data through a multi-view underwater camera system and generate three-dimensional point cloud data of the fish body based on the principle of binocular vision, from which growth characteristics and behavioral characteristics data are extracted.

[0110] The stereo vision recognition module is a data acquisition and preprocessing unit for a stereo vision-based fish grouping and directional feeding system in net cages. Its hardware interface is connected to a multi-view underwater camera system.

[0111] The stereo vision recognition module is configured to receive the raw video stream output by the multi-view underwater camera system, execute the point cloud reconstruction and feature extraction algorithms defined in the previous method steps, and send the generated growth feature and behavioral feature data to the intelligent decision module through the data bus.

[0112] The intelligent decision-making module is configured to divide the target population based on the feature data using a clustering algorithm, and generate differentiated feeding strategies based on growth characteristics and environmental parameters.

[0113] The intelligent decision-making module is the core computing unit of the cage fish grouping and directional feeding system based on stereo vision. Its input end is connected to the output end of the stereo vision recognition module and the environmental sensor interface.

[0114] The intelligent decision-making module is configured to call the clustering algorithm defined in the previous steps to process the input feature data, generate the target group segmentation results, and combine them with real-time environmental parameters to generate differentiated feeding strategies. Finally, the strategy data is output to the group feeding execution module.

[0115] The group feeding execution module includes a multi-outlet directional feeder configured to deliver feed to the target area according to the differentiated feeding strategy.

[0116] The group feeding execution module is the action execution unit of a cage-culture fish grouping and directional feeding system based on stereoscopic vision. Its control input is connected to the output of the intelligent decision module, and its mechanical output is connected to the multi-outlet directional feeder. The group feeding execution module is configured to parse the received differentiated feeding strategy and drive the multi-outlet directional feeder to perform physical delivery actions according to the feeding amount and feeding area specified in the strategy.

[0117] The central processing unit is used to schedule data interaction and instruction execution among various modules.

[0118] The central processing unit is the main control hardware of the cage fish grouping and directional feeding system based on stereo vision. It establishes communication links with the stereo vision recognition module, the intelligent decision-making module, and the grouping and feeding execution module through the system bus.

[0119] The central processing unit is configured to manage the timing of data interactions between modules, distribute synchronization clock signals, and monitor the status of each module to coordinate the global process of instruction execution.

[0120] Specifically, the intelligent decision-making module includes a nutrient requirement analysis unit, a feed adaptation unit, and an instruction generation unit, which are respectively configured to calculate the target feeding amount, match the feed nutrient ratio, and generate standardized feeding instructions containing water layer distribution indicators.

[0121] The internal software architecture of the intelligent decision-making module is further divided into three logical sub-units: the nutrition requirement analysis unit, the feed adaptation unit, and the instruction generation unit.

[0122] The nutritional requirements analysis unit receives cluster feature data at its input end and is connected to the instruction generation unit at its output end to execute the target feeding amount calculation logic defined above.

[0123] The feed adaptation unit receives environmental parameter data at its input end and is connected to the instruction generation unit at its output end to execute the feed matching logic defined above.

[0124] The instruction generation unit summarizes the output results of the above two units, encapsulates them to generate standardized feeding instructions containing water layer distribution identifiers, and outputs them to downstream modules.

[0125] Specifically, the feature is that the group feeding execution module further includes a group feeding controller, configured to parse the water layer distribution marker, generate a first control signal to drive the feeder to execute a direct-fall feeding mode when the marker indicates the bottom layer, and generate a second control signal to drive the feeder to execute a wind-driven feeding mode when the marker indicates the middle and upper layers.

[0126] The hardware circuit of the group feeding execution module also integrates a group feeding controller, whose signal input terminal is connected to the intelligent decision module to receive standardized feeding instructions.

[0127] The group feeding controller is configured to parse the water layer distribution identifier in the command: If the water layer distribution indicator points to the bottom layer, the group feeding controller outputs the first control signal to the drive circuit of the multi-outlet directional feeder to trigger the direct-fall feeding mode defined above. If the water layer distribution indicator points to the upper layer, the group feeding controller outputs a second control signal to the drive circuit of the multi-outlet directional feeder to trigger the wind-driven feeding mode defined above.

[0128] In one specific embodiment, a large-scale deep-sea cage aquaculture base cultivates large yellow croaker. The system deployment environment includes: a multi-lens underwater camera system (Hikvision DS-2CD6xxx series) installed at the four corners of the cages; a central processing unit (configured with an Intel Core i7 processor and 32GB of RAM) deployed in a shore-based equipment room; and a multi-outlet directional feeder (model: Custom-Feeder-V2, with four independent feed outlets) suspended above the cages. This embodiment aims to demonstrate the specific operation process of a cage-based fish grouping and directional feeding system based on stereoscopic vision during the high-temperature summer season (water temperature 26°C).

[0129] Step S1: Data acquisition and feature extraction.

[0130] The stereo vision recognition module is activated, driving the multi-view underwater camera system to acquire video data from inside the fish cage at a frame rate of 30fps. The stereo vision recognition module calls the built-in binocular vision algorithm, reconstructing the video data into three-dimensional point cloud data of the fish based on the principle of binocular vision.

[0131] In this embodiment, the point cloud density of the fish's 3D point cloud data is set to 5000 points per cubic meter, and the typical depth measurement accuracy is 2mm (according to the manufacturer's technical documentation). Subsequently, the stereo vision recognition module extracts the growth characteristics (including estimated body length and weight) and behavioral characteristic data (including swimming speed and aggregation degree) of 2000 large yellow croakers in the net cage from the fish's 3D point cloud data, forming a feature data set, which is then sent to the intelligent decision-making module via gigabit Ethernet.

[0132] Step S2: Group segmentation and strategy generation.

[0133] The intelligent decision-making module receives feature data and calls the K-means clustering algorithm (with the number of clusters set to 3, based on engineering experience values ​​typical of large yellow croaker body size differentiation in this scenario) to divide the 2000 fish into 3 target groups: Target group: average weight 50g, distributed in the upper layer of the net cage; Target group: average weight 150g, distributed in the middle layer of the net cage; Target group: average weight 300g, distributed at the bottom of the net cage.

[0134] The nutrient demand analysis unit inside the intelligent decision-making module reads the current water temperature of 26°C and the seasonal rhythm (value 1.0 in summer), and calculates the target feeding amount for each target group based on the growth characteristics of each target group: kg, kg, kg.

[0135] Meanwhile, the feed adaptation unit analyzed the matching degree between the current dissolved oxygen level (6.5 mg / L) and the nutritional components of the candidate feeds, and selected the feed code with a protein content of 45% as the feed type.

[0136] Finally, the instruction generation unit generates water layer distribution identifiers based on the depth distribution of each population: for and , it sets (upper middle layer); for , it sets (lower layer). The above data is encapsulated into 3 standardized feeding instructions and sent to the group feeding execution module.

[0137] Step S3: Pattern parsing and execution control.

[0138] The group feeding controller receives standardized feeding instructions and parses the water layer distribution identifiers within them.

[0139] For the target group: Since the water layer distribution indicator is equal to 2, the group feeding controller generates a second control signal. This signal drives the multi-outlet directional feeder to start the high-pressure blower to the air pressure Pa (typical value) and open the airflow regulating valve to the specified opening degree, putting the system into the air-assisted feeding mode. In this mode, a total of 7.5 kg of feed is blown by the airflow to the upper layer of the cage.

[0140] For the target group: Since the water layer distribution indicator is equal to 1, the group feeding controller generates the first control signal. This signal shuts down the high-pressure blower and adjusts the flow regulating valve to its opening position, putting the system into a direct-fall feeding mode. In this mode, 9.0 kg of feed falls freely along the vertical feeding channel to the bottom of the cage.

[0141] Step S4: Feedback and Model Update.

[0142] Twenty-four hours after feeding, the stereoscopic vision recognition module collected data again and calculated the actual growth rate of each group. The system compared this data with the expected growth rate predicted by the nutritional requirements analysis unit the previous day and calculated the deviation.

[0143] In this embodiment, it is assumed that the population's g / day is greater than a preset threshold (set at 0.05 g / day, based on experience from high-density aquaculture projects). At this time, the intelligent decision-making module automatically triggers the parameter adjustment logic, uses the gradient descent method to correct the temperature correction factor in the feeding parameters, and writes the updated parameters into the nutrient requirement model, completing one closed-loop optimization.

[0144] Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for grouping and targeted feeding of fish in net cages based on stereoscopic vision, characterized in that, Includes the following steps: The system collects video data of fish schools using a multi-view underwater camera system, and processes the video data based on the principle of binocular vision to generate three-dimensional point cloud data of the fish bodies. The body length and weight data of each individual fish are extracted from the three-dimensional point cloud data of the fish body as growth characteristics. Combined with behavioral characteristic data, a clustering algorithm is used to divide the fish group into at least two target groups with different growth characteristics. Based on the growth characteristics of each target population and current environmental parameters, generate differentiated feeding strategies that include feed type, feeding amount and feeding area; The multi-outlet directional feeder is controlled to accurately deliver feed to the activity areas of each target group according to the differentiated feeding strategy.

2. The method for grouping and targeted feeding of cage-cultured fish based on stereoscopic vision according to claim 1, characterized in that, The combined behavioral feature data specifically includes: A deep learning model is used to identify the biological characteristics of fish, and an inter-frame matching algorithm is used to continuously track each individual fish in consecutive video frames to obtain swimming speed and feeding frequency as behavioral characteristic data.

3. The method for grouping and targeted feeding of cage-cultured fish based on stereoscopic vision according to claim 1, characterized in that, The method of using clustering algorithms to divide a school of fish into at least two target groups with different growth characteristics includes any of the following: Method 1: Calculate the average body length and divide fish individuals whose actual body length is greater than, within, or less than the preset proportion of the average body length into fast, normal, or slow growth groups respectively; Method 2: Using body length as the core feature, a multi-dimensional feature space is constructed by combining water temperature, dissolved oxygen, salinity, pH value and flow rate. The K-means clustering algorithm is used to dynamically divide the fish population into groups of a preset number or groups whose number is automatically determined based on the clustering effectiveness index.

4. The method for grouping and targeted feeding of cage-cultured fish based on stereoscopic vision as described in claim 1, characterized in that, The generation of differentiated feeding strategies specifically includes: The nutritional requirement model, trained based on fish species data, grouping characteristics, and seasonal rhythms, is used to calculate the target feeding amount for each target group. The fitness analysis model is used to quantitatively analyze the matching relationship between environmental parameters and feed nutrients, output feed fitness scores and match feeds with corresponding nutrient ratios. After performing a logical consistency check on the feeding strategy data, a standardized feeding instruction containing water layer distribution identifiers is generated.

5. The method for grouping and targeted feeding of cage-cultured fish based on stereoscopic vision according to claim 4, characterized in that, The controlled multi-outlet directional feeder, according to the differentiated feeding strategy, precisely delivers feed to the activity areas of each target group, specifically including: Parse the water layer distribution identifier in the standardized feeding instructions; If the identifier indicates that the target group is located at the bottom of the cage, then the feeder is controlled to execute a direct-fall feeding mode; If the identifier indicates that the target group is distributed in the upper layer of the cage, then the feeder is controlled to execute the wind-driven feeding mode.

6. The method for grouping and targeted feeding of cage-cultured fish based on stereoscopic vision according to claim 5, characterized in that, The direct-fall feeding mode includes adjusting the flow regulating valve of the vertical feeding channel to control the feeding rate; The air-feeding mode includes adjusting the air pressure of the high-pressure blower and the airflow regulating valve to control the airflow intensity.

7. The method for grouping and targeted feeding of cage-cultured fish based on stereoscopic vision according to claim 1, characterized in that, It also includes feedback and adjustment steps: Periodically collect growth data and feed residue after feeding, and compare the actual and expected growth rates; If the deviation is lower than the preset threshold, the feeding parameters are adjusted and the nutritional requirement model is updated.

8. A system for grouping and targeted feeding of fish in net cages based on stereoscopic vision, characterized in that, include: The stereo vision recognition module is configured to acquire video data through a multi-view underwater camera system and generate three-dimensional point cloud data of the fish body based on the principle of binocular vision, from which growth characteristics and behavioral characteristics data are extracted; The intelligent decision-making module is configured to divide the target population based on the feature data using a clustering algorithm, and generate differentiated feeding strategies based on growth characteristics and environmental parameters. The group feeding execution module includes a multi-outlet directional feeder, configured to deliver feed to the target area according to the differentiated feeding strategy; The central processing unit is used to schedule data interaction and instruction execution among various modules.

9. The cage-cultured fish grouping and targeted feeding system based on stereoscopic vision according to claim 8, characterized in that, The intelligent decision-making module includes a nutrient requirement analysis unit, a feed adaptation unit, and an instruction generation unit, which are respectively configured to calculate the target feeding amount, match the feed nutrient ratio, and generate standardized feeding instructions containing water layer distribution indicators.

10. The cage fish grouping and targeted feeding system based on stereoscopic vision according to claim 9, characterized in that, The group feeding execution module also includes a group feeding controller, configured to parse the water layer distribution markers, generate a first control signal to drive the feeder to execute a direct-fall feeding mode when the marker indicates the bottom layer, and generate a second control signal to drive the feeder to execute a wind-driven feeding mode when the marker indicates the middle or upper layer.

Images