Intelligent isolation device and early warning method for same-species predation of carnivorous fish based on machine vision

By monitoring fish posture using machine vision and generating a non-contact bubble barrier, the problem of cannibalism among carnivorous fish has been solved, improving the survival rate of seedlings and the level of intelligent aquaculture management.

CN121730234BActive Publication Date: 2026-05-08SICHUAN AGRI UNIV
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
CN202610226341.3
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2026-02-26
Publication Date
2026-05-08
Estimated Expiration
2046-02-26

AI Technical Summary

Technical Problem

Existing technologies cannot effectively intervene in the moment when carnivorous fish are cannibalistic, and traditional methods are prone to causing fish injury and stress response. Existing video monitoring systems lack real-time blocking means, have a high false alarm rate, and cannot meet millisecond-level requirements.

Method used

A machine vision-based intelligent isolation device is used to monitor the fish's posture in real time through a suspended camera module and an array-type air curtain generator. A non-contact bubble barrier is generated by an array of microporous aeration pipes controlled by an electromagnetic valve to block potential cannibalistic behavior.

Benefits of technology

This has resulted in a significantly improved survival rate for carnivorous fish fry, avoiding mechanical damage and stress, and enhancing the level of intelligent aquaculture management.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN121730234B_ABST
    Figure CN121730234B_ABST
Patent Text Reader

Abstract

The present application relates to the field of fish farming early warning, in particular to a kind of intelligent isolation device and early warning method for carnivorous fish cannibalism based on machine vision.The device mainly includes a breeding tank, a visual acquisition module, a system control and data processing mechanism, and an array air curtain generating mechanism at the bottom of the tank.The visual acquisition module acquires real-time water images in the breeding tank.A posture perception network capable of outputting target spatial position and multiple structural key points is used in embedded terminal operation.The vector angle and Euclidean distance of fish spine are calculated in real time by extracting anatomical key points such as fish head, dorsal fin and tail handle.When the target fish presents a high-curvature attack posture, the system immediately locks the conflict area through coordinate mapping and drives the integrated micro-porous aeration pipe array to release compressed air, generating a vertical non-contact bubble barrier between the attacker and the victim, achieving physical isolation.The present application is suitable for fish farming.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of fish farming early warning, specifically to an intelligent isolation device and early warning method for cannibalism among carnivorous fish based on machine vision. Background Technology

[0002] As aquaculture develops towards intensification and factory farming, high-density recirculating aquaculture systems (RAS) have become the mainstream model. During the fry rearing stage of high-value carnivorous fish such as grouper, mandarin fish, largemouth bass, and fighting fish, cannibalism is highly likely due to their natural aggression and territoriality. Studies have shown that during the metamorphosis and growth differentiation stages of fry, cannibalism can cause losses as high as 30%-50%, severely restricting aquaculture efficiency.

[0003] To solve this problem, existing technical solutions mainly fall into two categories:

[0004] The first type is periodic mechanical grading. This involves periodically separating fish fry of different sizes using manual netting or sieves. However, this method has significant drawbacks: first, it is reactive and cannot intervene immediately when aggression occurs, only providing remedial action "after the fact"; second, mechanical grading is a contact operation, which can easily cause mechanical damage to the fish's surface, such as scale loss, leading to secondary infections like saprolegniasis; and third, frequent human interference can cause strong stress responses in the fish, affecting their feeding and growth.

[0005] The second category is early warning systems based on traditional video surveillance. Currently, there are some monitoring devices for underwater environments on the market. While they can use image processing technology to identify the number of fish or their approximate movement trajectories, they generally suffer from the following shortcomings: First, most algorithms only remain at the "monitoring" level, lacking linkage with actuators and failing to achieve closed-loop control; second, existing behavior recognition algorithms are mostly based on simple movement speed thresholds, making it difficult to distinguish between normal rapid cruising and the charging posture before an attack, resulting in a high false alarm rate; finally, there is a lack of effective non-contact blocking methods. Even if an attack is detected, it usually only requires manual intervention through alarm prompts, which is insufficient to meet the millisecond-level blocking requirements.

[0006] To adapt to underwater vision, fish eyes have evolved a fixed spherical lens and a wide field of view, making it difficult for them to focus quickly and filter out dynamic blur. A bubble curtain creates a dense and irregular air-water interface, forming a constantly shifting and breaking optical noise barrier in the water. This continuous light scattering and image distortion directly attack the natural weaknesses of fish vision, preventing them from consistently acquiring clear images. This leads to sensory interference and behavioral obstacles, potentially affecting their hunting or communication with other fish.

[0007] In conclusion, developing an intelligent device capable of accurately identifying pre-attack posture using machine vision and achieving millisecond-level instantaneous blocking through non-contact methods such as pneumatic interference would have significant application value for improving the survival rate of carnivorous fish fry and animal welfare. Summary of the Invention

[0008] The purpose of this invention is to overcome the shortcomings of the prior art and provide a machine vision-based intelligent isolation device and early warning method for cannibalism among carnivorous fish. This solves the problem of cannibalism among valuable carnivorous fish fry in high-density aquaculture, avoids stress damage to fish caused by traditional mechanical driving, and significantly improves the survival rate of fry and the level of intelligence in aquaculture management.

[0009] The present invention achieves the above-mentioned objectives by adopting the following technical solution: Firstly, the present invention provides a machine vision-based intelligent isolation device for cannibalism in carnivorous fish, comprising:

[0010] The breeding box 3, the visual acquisition module, the system control and data processing mechanism, and the array-type air curtain generating mechanism are included. The system control and data processing mechanism is a chassis-embedded edge computing terminal 1. The breeding box 3 is located above the chassis-embedded edge computing terminal 1. The visual acquisition module consists of a suspended main camera module 8 and an auxiliary camera module 6. The suspended main camera module 8 is connected to the support frame 4 on both sides through flexible flat cables 7 on both sides. The auxiliary camera module 6 is distributed on the four sides of the inner wall of the breeding box 3. The array-type air curtain generating mechanism consists of an integrated microporous aeration pipe array 5, which is controlled by a solenoid valve and is regularly distributed at the bottom of the breeding box 3.

[0011] A visual acquisition module, consisting of a suspended main camera module and an auxiliary camera module, acquires real-time video streams of the water in the aquaculture tank, monitoring the posture of all fish. Through a posture perception network running in an embedded edge computing terminal, the anatomical points of the fish's head, body center, dorsal fin, and tail are extracted, and the vector angle and Euclidean distance of the fish's spine are calculated in real time. When the embedded edge computing terminal detects that the target fish is exhibiting a high-curvature attack posture, indicating a possible cannibalistic event, the system control and data processing mechanism immediately locks the conflict area through coordinate mapping and drives the integrated microporous aeration tube array at the bottom of the aquaculture tank to respond within milliseconds, releasing compressed air to generate a vertical, non-contact bubble barrier between the attacker and the attacked, achieving physical blocking.

[0012] Furthermore, a guide rail assembly 2 is installed on the upper part of the chassis-embedded edge computing terminal 1, distributed on both sides of the breeding box 3. The guide rail assembly 2 consists of a guide rail 201 and a limiting block 202. The support frame 4 contacts the guide rail 201 through a pulley in the bottom hole.

[0013] Furthermore, the suspended main camera module 8 consists of a pod 801 and a main camera 802. The pod 801 is connected to the flexible flat cables 7 on both sides and the wheels in the support frames 4 on both sides. The main camera 802 is located below the pod 801.

[0014] Furthermore, the integrated microporous aeration tube array 5 is divided into four-hole microporous aeration tube 501, three-hole microporous aeration tube 502 and single-hole microporous aeration tube 503 according to its location, and is driven by a solenoid valve.

[0015] Secondly, the present invention provides a machine vision-based intelligent early warning method for cannibalism in carnivorous fish, applied to the aforementioned machine vision-based intelligent isolation device for cannibalism in carnivorous fish. The early warning method includes:

[0016] S1. The video stream of the aquaculture tank acquired by the visual acquisition module is preprocessed. Gaussian filtering is used to remove salt and pepper noise caused by water bubbles and suspended particles, and adaptive histogram equalization is used to enhance image contrast.

[0017] S2. Input the preprocessed image into the pose perception network to extract the anatomical points of the fish's head, body center, dorsal fin, and tail.

[0018] The pose perception network employs a pre-trained YOLOv8-Pose neural network model. The YOLOv8-Pose neural network model uses a hierarchical visual perception architecture: the backbone network uses a CSPDarknet structure for feature extraction, transforming RGB pixel values ​​layer by layer into the fish outline, fish texture, morphological differences between the head and tail, and overall structural features of the curved posture; the head network is configured to output two types of information for each detected target.

[0019] First type of bounding box information:

[0020] ;

[0021] In the formula, B represents the bounding box itself. The x-axis coordinates of the center point of the bounding box. The y-coordinate of the center point of the bounding box. Indicates the width of the bounding box. Indicates the height of the bounding box;

[0022] Bounding boxes are used to define the spatial extent of the target fish and serve as object-level constraints for point extraction, pose analysis, and behavior determination.

[0023] The second type is the coordinates of the fish's body position:

[0024] Identify and record the spatial coordinates of the head, body center or dorsal fin, and tail of the fish skeleton;

[0025] Head coordinates: ;

[0026] Body center or dorsal fin coordinates: ;

[0027] Tail coordinates: ;

[0028] S3. For each detected fish target, calculate the spinal curvature angle based on the fish's position coordinates;

[0029] First, calculate the corresponding vector based on the coordinates of the fish's body position;

[0030] Vector from the center of the body or dorsal fin to the head:

[0031] ;

[0032] Vector from the center of the body or dorsal fin to the tail:

[0033] ;

[0034] Then, the spinal curvature angle is calculated using the vector dot product formula. :

[0035] , ;

[0036] S4. If the spinal curvature angle of an individual is in multiple consecutive frames Less than the threshold Then find the fish closest to the individual, and record the coordinates of the closest point of this fish to the fish in the attack charging state, and calculate the Euclidean distance:

[0037] ;

[0038] In the formula, Represents Euclidean distance. , These are the coordinates of the individual fish in its attack charging state. , These are the coordinates of the fish closest to the former;

[0039] S5. Define the attack detection logic function. :

[0040] ;

[0041] In the formula, This indicates a preset Euclidean distance threshold;

[0042] like If the value is 0, it is determined that no cannibalistic event of the same kind has occurred;

[0043] like If so, it is determined that a cannibalistic event of the same kind has occurred.

[0044] Furthermore, this early warning method also includes:

[0045] When a possible cannibalistic event is determined, the coordinate mapping of the event's occurrence is calculated:

[0046] ;

[0047] These are the image coordinates in the video stream captured by the camera. These are the actual coordinates, i.e., the coordinates of the corresponding integrated microporous aeration tube array;

[0048] .

[0049] Furthermore, this early warning method also includes:

[0050] When a cannibalistic event is detected, the system control and data processing mechanism sends a command to control the integrated microporous aeration tube array located at the bottom of the breeding tank to release compressed air, creating a vertical non-contact bubble barrier between the attacker and the attacked, thereby achieving physical blocking.

[0051] The beneficial effects of this invention are as follows:

[0052] In terms of software, this invention abandons traditional motion trajectory tracking and adopts a posture perception network that can simultaneously output the target's spatial position and multiple structural key points. By extracting key points from the fish's head, dorsal fin, and caudal peduncle, it calculates the "spinal curvature" and defines an attack determination logic function based on a distance threshold. This is used to determine whether a fish is adopting an aggressive posture. When a fish is detected adopting a "C-shaped" or "S-shaped" posture to prepare for an attack, it is determined to be a possible cannibalistic event, achieving a technological leap from "post-event alarm" to "pre-event warning".

[0053] In terms of hardware, this invention designs a bottom-matrix air curtain generating mechanism, which includes an integrated microporous aeration tube array controlled by solenoid valves. Visual coordinates are mapped to physical pores using a homography matrix. Once a potential cannibalistic event is detected, the corresponding micro-solenoid valves in the relevant area instantly open, generating a vertical bubble wall to separate the two fish. This non-contact isolation effectively avoids stress and injury to the fish.

[0054] This invention effectively solves the problem of cannibalism among valuable carnivorous fish fry in high-density aquaculture by using a hardware-software synergy mechanism of "visual early warning + air curtain isolation". It avoids stress damage to fish caused by traditional mechanical driving and significantly improves the survival rate of fry and the level of intelligence in aquaculture management. Attached Figure Description

[0055] Figure 1 This is a flowchart of an intelligent early warning method for cannibalism in carnivorous fish based on machine vision, provided by the present invention.

[0056] Figure 2 This is a schematic diagram of the intelligent isolation device provided by the present invention;

[0057] Figure 3 This is a front view of the intelligent isolation device provided by the present invention;

[0058] Figure 4 This is a side view of the intelligent isolation device provided by the present invention;

[0059] Figure 5 This is a top view schematic diagram of the intelligent isolation device provided by the present invention;

[0060] Figure 6 This is a schematic diagram of the structure of the suspended main camera module provided by the present invention;

[0061] Figure 7 This is a schematic diagram of the integrated microporous aeration tube array provided by the present invention;

[0062] Figure 8 This is a schematic diagram of the structure of the auxiliary camera module provided by the present invention.

[0063] In the attached diagram, 1 is the chassis-embedded edge computing terminal, 2 is the guide rail assembly, 201 is the guide rail, 202 is the limiting block, 3 is the breeding box, 4 is the support frame, 5 is the integrated microporous aeration tube array, 501 is the four-hole microporous aeration tube, 502 is the three-hole microporous aeration tube, 503 is the single-hole microporous aeration tube, 6 is the auxiliary camera module, 7 is the flexible flat cable, 8 is the suspended main camera module, 801 is the pod, and 802 is the main camera. Detailed Implementation

[0064] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings.

[0065] Reference Figures 2 to 5As shown, this invention provides an intelligent isolation device for cannibalism in carnivorous fish based on machine vision, comprising: a rearing tank 3, a visual acquisition module, a system control and data processing mechanism, and an array-type air curtain generating mechanism. The system control and data processing mechanism is a chassis-embedded edge computing terminal 1. The rearing tank 3 is located above the chassis-embedded edge computing terminal 1. The visual acquisition module consists of a suspended main camera module 8 and an auxiliary camera module 6. The suspended main camera module 8 is connected to the support frames 4 on both sides via flexible flat cables 7 on both sides. The auxiliary camera modules 6 are distributed on the four sides of the inner wall of the rearing tank 3. The array-type air curtain generating mechanism consists of an integrated microporous aeration pipe array 5, controlled by a solenoid valve, and regularly distributed at the bottom of the rearing tank 3.

[0066] A visual acquisition module, consisting of a suspended main camera module and an auxiliary camera module, acquires real-time video streams of the water in the aquaculture tank, monitoring the posture of all fish. Through a posture perception network running in an embedded edge computing terminal, key anatomical points of the fish's head, body center, dorsal fin, and tail are extracted. The vector angle and Euclidean distance of the fish's spine are calculated in real time. When the embedded edge computing terminal detects that the target fish is exhibiting a high-curvature attack posture, indicating a possible cannibalistic event, the system control and data processing mechanism immediately locks the conflict area through coordinate mapping. It then drives the integrated microporous aeration tube array at the bottom of the aquaculture tank to respond within milliseconds, releasing compressed air to generate a vertical, non-contact bubble barrier between the attacker and the attacked, achieving physical blocking.

[0067] Reference Figure 4 and Figure 5 As shown, a guide rail assembly 2 is installed on the upper part of the chassis-embedded edge computing terminal 1, distributed on both sides of the breeding box 3. The guide rail assembly 2 consists of a guide rail 201 and a limiting block 202. The support frame 4 contacts the guide rail 201 through pulleys in the bottom holes. Driven by the motor, it can slide along the horizontal plane of the guide rail 201. The limiting block 202 serves to restrict and protect the movement of the support frame within a limited range.

[0068] Reference Figure 6 As shown, the suspended main camera module 8 consists of a pod 801 and a main camera 802. The pod 801 is connected to flexible flat cables 7 on both sides, which are connected to wheels in the support frames 4 on both sides. The main camera 802 is located below the pod 801. The flexible flat cables 7 on both sides of the pod 801 are connected to wheels in the support frames 4 on both sides. Through the coordinated rotation of the wheels, the flexible flat cables 7 on both sides can be extended and retracted, allowing the main camera 802 to move horizontally and vertically along the flexible flat cables.

[0069] Reference Figure 6 and Figure 8As shown, under the combined effect of the above two points, the main camera module 8 can move up, down, left, right, and forward and backward within the spatial range under the control of the control system, and with the assistance of the auxiliary camera module 6, it can achieve all-round monitoring of the water body and the position and posture of all individual fish.

[0070] Reference Figure 7 As shown, the integrated microporous aeration tube array 5 is divided into four-hole microporous aeration tube 501, three-hole microporous aeration tube 502 and single-hole microporous aeration tube 503 according to its location. It is driven by a solenoid valve and can accurately release compressed air at the corresponding coordinate after receiving the system action command.

[0071] Reference Figure 1 As shown, this invention provides a machine vision-based intelligent early warning method for cannibalism in carnivorous fish, comprising:

[0072] S1. The input visual acquisition module acquires video streams from inside the breeding tank in real time at a frame rate of 30fps and performs preprocessing: Gaussian filtering is used to remove salt and pepper noise caused by water bubbles and suspended particles; adaptive histogram equalization is used to enhance image contrast and improve the distinguishability of fish edge features.

[0073] S2. Input the preprocessed image into the chassis-embedded edge computing terminal and run a pose perception network capable of simultaneously outputting the target's spatial location and multiple structural key points, such as the pre-trained YOLOv8-Pose neural network model. This network employs a hierarchical visual perception architecture: the backbone uses a CSPDarknet structure for feature extraction, transforming RGB pixel values ​​layer by layer into the fish's outline, texture, morphological differences between the head and tail, and overall structural features of the bending posture; the head is configured to output two types of information for each detected target:

[0074] Bounding box information is used to locate the spatial position of the fish:

[0075] ;

[0076] In the formula, B represents the bounding box itself. The x-axis (horizontal direction) coordinates represent the center point of the bounding box. The y-axis (vertical direction) coordinates representing the center point of the bounding box. Indicates the width of the bounding box. This indicates the height of the bounding box.

[0077] Key coordinates of the fish body:

[0078] The model identified three key points of the fish skeleton: the head ( ), body center / dorsal fin ( ), tail ( The spatial coordinates at this point are recorded as the reference for attitude analysis.

[0079] Head coordinates: ;

[0080] Body center / dorsal fin coordinates: ;

[0081] Tail coordinates: ;

[0082] Obtain the key coordinates of the fish body:

[0083] ;

[0084] S3. For each detected fish target, calculate the spinal curvature angle formed by the above three key points. The specific calculation is achieved using the vector dot product formula:

[0085] Define two vectors:

[0086] Body center / dorsal fin ( ) to head ( Direction vector:

[0087] ;

[0088] Body center / dorsal fin ( ) to the tail ( Direction vector:

[0089] ;

[0090] The angle of the fish's spine for:

[0091] ;

[0092] ;

[0093] If the fish swims normally ≈180°. If the fish is in an attack-gathering state, its body will bend noticeably, forming a "C" or "S" shape, and the angle θ between the fish's bodies will decrease significantly, i.e., less than the threshold. (Preset threshold:) =150°); at the same time, the system further analyzes the changing trend of the spinal curvature angle in multiple consecutive frames to distinguish between instantaneous turning behavior and continuous attack charging behavior.

[0094] S4. If the spinal curvature angle of an individual is in multiple consecutive frames Less than the threshold Find the nearest fish individual and record the coordinates of the closest point to the fish in the attack charging state. Calculate the Euclidean distance. :

[0095] ;

[0096] In the formula, , These are the coordinates of the individual fish in its attack charging state. , It is the coordinates of the fish closest to the former, with a predefined Euclidean distance threshold. 10cm;

[0097] S5. Define the attack detection logic function. :

[0098] ;

[0099] like If the value is 0, it is determined that no cannibalistic event of the same kind has occurred;

[0100] like If so, it is determined that a possible cannibalistic event has occurred;

[0101] When a possible cannibalistic event is determined, the coordinate mapping of the event's occurrence is calculated:

[0102] ;

[0103] These are the image coordinates in the video stream captured by the camera. These are the actual coordinates, i.e., the coordinates of the corresponding integrated microporous aeration tube array;

[0104] .

[0105] Its function is to describe the geometric relationship between the "plane seen by the camera" and the "real physical plane". It is obtained through calibration during system initialization. The above formula can convert the attack behavior detection results based on the image pixel coordinate system into the corresponding position in the physical execution coordinate system, thereby realizing precise control of the integrated microporous aeration tube array.

[0106] When a cannibalistic event is detected, the system control and data processing mechanism sends a command to control the integrated microporous aeration tube array located at the bottom of the breeding tank to release compressed air, creating a vertical non-contact bubble barrier between the attacker and the attacked, thereby achieving physical blocking.

[0107] The above description is merely a preferred embodiment of the present invention. It should be understood that the present invention is not limited to the forms disclosed herein and should not be construed as excluding other embodiments. It can be used in various other combinations, modifications, and environments, and can be altered within the scope of the concept described herein through the above teachings or related technologies or knowledge. Modifications and variations made by those skilled in the art that do not depart from the spirit and scope of the present invention should be within the protection scope of the appended claims.

Claims

1. A machine vision-based intelligent isolation device for cannibalism in carnivorous fish, characterized in that, include: The breeding box (3), visual acquisition module, system control and data processing mechanism and array-type air curtain generating mechanism, wherein the system control and data processing mechanism is a chassis-embedded edge computing terminal (1), the breeding box (3) is located above the chassis-embedded edge computing terminal (1), the visual acquisition module consists of a suspended main camera module (8) and an auxiliary camera module (6), the suspended main camera module (8) is connected to the support frame (4) on both sides through flexible flat cables (7) on both sides, the auxiliary camera module (6) is distributed on the four sides of the inner wall of the breeding box (3), and the array-type air curtain generating mechanism consists of an integrated microporous aeration pipe array (5), controlled by a solenoid valve, and regularly distributed at the bottom of the breeding box (3); A visual acquisition module, consisting of a suspended main camera module and an auxiliary camera module, acquires real-time video streams of the water in the aquaculture tank, monitoring the posture of all fish. Through a posture perception network running in an embedded edge computing terminal, the anatomical points of the fish's head, body center, dorsal fin, and tail are extracted, and the vector angle and Euclidean distance of the fish's spine are calculated in real time. When the embedded edge computing terminal detects that the target fish is exhibiting a high-curvature attack posture, i.e., a cannibalistic event has occurred, the system control and data processing mechanism immediately locks the conflict area through coordinate mapping, and drives the integrated microporous aeration tube array at the bottom of the aquaculture tank to respond within milliseconds, releasing compressed air to generate a vertical non-contact bubble barrier between the attacker and the attacked, achieving physical blocking.

2. The intelligent isolation device for cannibalism among carnivorous fish based on machine vision according to claim 1, characterized in that, The chassis-embedded edge computing terminal (1) is equipped with a guide rail assembly (2) on the upper part, which is distributed on both sides of the breeding box (3). The guide rail assembly (2) consists of a guide rail (201) and a limiting block (202). The support frame (4) is in contact with the guide rail (201) through a pulley in the bottom hole.

3. The intelligent isolation device for cannibalism among carnivorous fish based on machine vision according to claim 1, characterized in that, The suspended main camera module (8) consists of a pod (801) and a main camera (802). The pod (801) is connected to the flexible flat cables (7) on both sides and the wheels in the support frames (4) on both sides. The main camera (802) is located below the pod (801).

4. The intelligent isolation device for cannibalism among carnivorous fish based on machine vision according to claim 1, characterized in that, The integrated microporous aeration tube array (5) is divided into four-hole microporous aeration tube (501), three-hole microporous aeration tube (502) and single-hole microporous aeration tube (503) according to its location, and is driven by a solenoid valve.

5. A machine vision-based intelligent early warning method for cannibalism in carnivorous fish, applied to the machine vision-based intelligent isolation device for cannibalism in carnivorous fish as described in any one of claims 1-4, characterized in that, The early warning method includes: S1. The video stream of the aquaculture tank acquired by the visual acquisition module is preprocessed. Gaussian filtering is used to remove salt and pepper noise caused by water bubbles and suspended particles, and adaptive histogram equalization is used to enhance image contrast. S2. Input the preprocessed image into the pose perception network to extract the anatomical points of the fish's head, body center, dorsal fin, and tail. The pose perception network employs a pre-trained YOLOv8-Pose neural network model. The YOLOv8-Pose neural network model uses a hierarchical visual perception architecture: the backbone network uses a CSPDarknet structure for feature extraction, transforming RGB pixel values ​​layer by layer into the fish outline, fish texture, morphological differences between the head and tail, and overall structural features of the curved posture; the head network is configured to output two types of information for each detected target. First type of bounding box information: ; In the formula, B represents the bounding box itself. The x-axis coordinates of the center point of the bounding box. The y-coordinate of the center point of the bounding box. Indicates the width of the bounding box. Indicates the height of the bounding box; Bounding boxes are used to define the spatial extent of the target fish and serve as object-level constraints for point extraction, pose analysis, and behavior determination. The second type is the coordinates of the fish's body position: Identify and record the spatial coordinates of the head, body center or dorsal fin, and tail of the fish skeleton; Head coordinates: ( ); Body center or dorsal fin coordinates: ( ); Tail coordinates: ( ); S3. For each detected fish target, calculate the spinal curvature angle based on the fish's position coordinates; First, calculate the corresponding vector based on the coordinates of the fish's body position; Vector from the center of the body or dorsal fin to the head: ; Vector from the center of the body or dorsal fin to the tail: ; Then, the spinal curvature angle is calculated using the vector dot product formula. : , ; S4. If the spinal curvature angle of an individual is in multiple consecutive frames Less than the threshold Then find the fish closest to the individual, and record the coordinates of the closest point of this fish to the fish in the attack charging state, and calculate the Euclidean distance: ; In the formula, Represents Euclidean distance. , This indicates the coordinates of an individual fish that is charging up its attack. , Represents the coordinates of the fish closest to the former; S5. Define the attack detection logic function. : ; In the formula, This indicates a preset Euclidean distance threshold; like If the value is 0, it is determined that no cannibalistic event of the same kind has occurred; like If so, it is determined that a cannibalistic event of the same kind has occurred.

6. The intelligent early warning method for cannibalism in carnivorous fish based on machine vision according to claim 5, characterized in that, This early warning method also includes: When a possible cannibalistic event is determined, the coordinate mapping of the event's occurrence is calculated: ; In the formula, These are the image coordinates in the video stream captured by the camera. These are the actual coordinates, i.e., the coordinates of the corresponding integrated microporous aeration tube array; 。 7. The intelligent early warning method for cannibalism in carnivorous fish based on machine vision according to claim 5, characterized in that, This early warning method also includes: When a cannibalistic event is detected, the system control and data processing mechanism sends a command to control the integrated microporous aeration tube array located at the bottom of the breeding tank to release compressed air, creating a vertical non-contact bubble barrier between the attacker and the attacked, thereby achieving physical blocking.

Citation Information

Patent Citations

  • Automatic feed feeding device for fry breeding and control system of automatic feed feeding device

    CN119325937A

  • Ornamental aquarium

    JP2007068455A