Aquaculture intelligent feeding system and method based on unmanned ship

Abstract

The invention provides an aquaculture intelligent feeding system and method based on an unmanned ship, and the method comprises the steps: controlling the unmanned ship to move along a cruise path of an aquaculture water area and hover at each feeding point, and collecting a water surface video stream of the unmanned ship at the feeding points through a ship-borne camera; ship body pose data of the unmanned ship are synchronously obtained, affine transformation compensation is carried out on the water surface video stream based on the ship body pose data, and a stable video stream is obtained; the method comprises the following steps: determining a feeding state label of aquaculture at a feeding point under each video frame in a stable video stream, determining a feeding response index of the aquaculture at the feeding point at the current moment based on all the feeding state labels, and triggering a feeding stopping instruction based on the feeding response index to control a shipborne feeding machine to interrupt feed feeding at the feeding point. And automatic feeding is carried out on each feeding point on the cruise path of the aquaculture water area. By adopting the scheme of the invention, intelligent feeding control based on aquatic product ingestion response can be realized by combining the motion compensation of the unmanned ship and the accurate recognition of the ingestion state.

Description

Technical Field

[0001] This application relates to the field of aquaculture technology, and more specifically, to an intelligent feeding system and method for aquaculture based on an unmanned vessel. Background Technology

[0002] Aquaculture is an important industry for ensuring a stable supply of aquatic products. The accuracy of the feeding process directly affects the breeding efficiency, the quality of aquatic products, and the aquatic ecological environment. With the penetration of intelligent technology into the agricultural field, the use of unmanned boats and other equipment to realize automated feeding in aquaculture waters has become an important development direction for reducing labor costs and improving the uniformity of feeding. Its core lies in realizing dynamic feeding control through real-time perception of the feeding status of aquatic products.

[0003] In existing technologies, aquaculture feeding largely relies on manual timed and quantitative operations or the use of automatic feeders with fixed parameters, lacking the perception and response to the real-time feeding status of aquatic organisms. Although some unmanned vessel-based feeding systems can move and feed along preset paths, the motion blur caused by the swaying of the vessel makes it difficult to accurately extract feeding characteristics. Furthermore, they cannot dynamically adjust the feeding strategy based on feeding activity, which can easily lead to excessive feed waste or insufficient feeding, affecting aquatic growth and polluting water quality. Therefore, how to combine the motion compensation of unmanned vessels with the accurate identification of feeding status to achieve intelligent feeding control based on the feeding response of aquatic organisms has become a challenge for the industry. Summary of the Invention

[0004] This application provides an intelligent feeding system and method for aquaculture based on unmanned vessels, which can combine motion compensation of unmanned vessels with accurate identification of feeding status to achieve intelligent feeding control based on aquatic feeding response.

[0005] In a first aspect, this application provides an intelligent feeding method for aquaculture based on an unmanned vessel, wherein automatic feeding control is performed according to the feeding status of the aquatic organisms, and the method includes: Control the unmanned vessel to move along the cruise path of the aquaculture area and hover at each feeding point to start feeding; For each feeding point along the patrol route of the aquaculture area, the unmanned vessel collects video streams of the water surface at the feeding point using shipborne cameras. The hull pose data of the unmanned vessel is acquired synchronously, and then the affine transformation compensation of the water surface video stream of the unmanned vessel at the feeding point is performed based on the hull pose data to obtain a stable video stream after eliminating motion blur. The feeding status label of the farmed aquatic organisms at the feeding point is determined in each video frame of the stable video stream. Then, based on all the feeding status labels, feature extraction is performed on the video frames in the stable video stream to obtain the feeding behavior features of the farmed aquatic organisms in all video frames. All feeding behavior characteristics are decoupled to determine the feeding response index of aquatic organisms at the feeding point at the current moment. Based on the feeding response index, a stop feeding command is triggered to control the shipborne feeder to interrupt the feed delivery at the feeding point, thereby automatically feeding each feeding point along the cruise path of the aquaculture area.

[0006] In some embodiments, controlling the unmanned vessel to move along the cruise path of the aquaculture area and hover at each feeding point to initiate feed dispensing specifically includes: A dynamic path calibration algorithm is used to generate a cruise path for the aquaculture area based on the electronic fence of the target aquaculture area and the preset coordinates of the feeding points. Control the unmanned vessel to move along the cruise path of the aquaculture area; For each feeding point along the cruise path of the aquaculture area, when the unmanned vessel moves along the cruise path of the aquaculture area and arrives at the feeding point, the dual-axis attitude stabilization system of the unmanned vessel is activated to achieve hovering through propeller differential adjustment. Once the unmanned vessel has hovered and stabilized, the onboard feeder is activated by triggering a signal to dispense feed at the preset maximum amount, thereby controlling the unmanned vessel to hover at each feeding point and begin dispensing feed.

[0007] In some embodiments, performing affine transformation compensation on the surface video stream of the unmanned vessel at the feeding point based on the hull pose data to obtain a stable video stream after motion blur elimination specifically includes: Construct an affine transformation matrix based on the hull pose data; Based on the affine transformation matrix, offset correction is performed on each video frame in the surface video stream of the unmanned vessel at the feeding point to obtain a corrected video stream. The corrected video stream is sharpened and enhanced based on the pose change rate of the unmanned vessel in the hull pose data to generate a stable video stream after motion blur is eliminated.

[0008] In some embodiments, determining the feeding status label of aquatic organisms at feeding points in each video frame of the stable video stream specifically includes: The feeding status of farmed aquatic organisms in the stable video stream before the unmanned vessel feeds the aquatic organisms is labeled as low-activity; The feeding status tags of aquatic organisms in the initial video frame of the unmanned vessel feeding in the stable video stream are marked as high-activity, thereby determining the feeding status tags of aquatic organisms at each feeding point in the stable video stream.

[0009] In some embodiments, feature extraction is performed on video frames in the stable video stream based on all feeding state labels to obtain feeding behavior features of farmed aquatic organisms in all video frames, specifically including: Multi-scale filtering is performed on video frames marked with feeding status tags in the stable video stream to obtain an effective set of texture-enhanced video frames. For each video frame in the set of valid video frames, extract the inverse difference moment of the gray-level co-occurrence matrix in the video frame; Determine the grayscale difference statistical entropy and histogram of video frames; Based on the inverse difference moment, the gray-level difference statistical entropy and the histogram, multi-dimensional texture features of video frames are generated, and then multi-dimensional texture features of each video frame in the effective video frame set are obtained. Scale calibration was performed on various multi-dimensional texture features to obtain the feeding behavior features of farmed aquatic products in all video frames.

[0010] In some embodiments, feature decoupling of all feeding behavior characteristics, and then determining the feeding response index of aquatic organisms at the feeding point at the current moment, specifically includes: Construct a feature orthogonality matrix based on all feeding behavior characteristics; Determine the proportion of orthogonal features in the feature orthogonality matrix; The proportion of high-activity frames in the stable video stream whose feeding status label is high-activity is determined based on all feeding status labels. The orthogonal feature ratio and the high-activity frame ratio are fused to generate the feeding response index of aquatic organisms at the feeding point at the current moment.

[0011] In some embodiments, triggering a stop-feeding command based on the feeding response index to control the shipborne feeder to interrupt feed delivery at feeding points, thereby automatically feeding each feeding point along the cruise path of the aquaculture area, specifically includes: When the feeding response index is lower than a preset response threshold, a stop feeding command is triggered; Based on the stop feeding command, the shipborne feeder is controlled to interrupt the feed delivery at the feeding point, thus completing the control of the current feeding point; After feeding at the current feeding point is completed, the unmanned vessel is controlled to proceed to the next feeding point along the cruise route of the aquaculture area, repeating the above steps to automatically feed each feeding point along the cruise route of the aquaculture area.

[0012] Secondly, this application provides an intelligent feeding system for aquaculture based on an unmanned vessel, the system comprising: The data acquisition module is used to control the unmanned vessel to move along the cruise path of the aquaculture area and hover at each feeding point to start feeding. The acquisition module is used to collect video streams of the unmanned vessel on the water surface at each feeding point along the cruise route of the aquaculture area using a ship-mounted camera. The processing module is used to synchronously acquire the hull pose data of the unmanned vessel, and then perform affine transformation compensation on the water surface video stream of the unmanned vessel at the feeding point based on the hull pose data to obtain a stable video stream after eliminating motion blur. The processing module is used to determine the feeding status label of the farmed aquatic organisms at the feeding point in each video frame of the stable video stream, and then extract features from the video frames in the stable video stream based on all the feeding status labels to obtain the feeding behavior features of the farmed aquatic organisms in all video frames. The execution module is used to decouple all feeding behavior characteristics, thereby determining the feeding response index of aquatic organisms at the feeding point at the current moment. Based on the feeding response index, a stop feeding command is triggered to control the shipborne feeder to interrupt the feed delivery at the feeding point, thereby automatically feeding each feeding point along the cruise path of the aquaculture area.

[0013] The technical solutions provided by the embodiments disclosed in this application have the following beneficial effects: The intelligent feeding system and method for aquaculture based on unmanned surface vessels (USVs) provided in this application first controls the USV to move along a cruise path in the aquaculture area and hover at each feeding point to start feeding. For each feeding point on the cruise path, the USV's surface video stream at the feeding point is collected by an onboard camera. Simultaneously, the USV's hull pose data is acquired, and then affine transformation compensation is performed on the surface video stream at the feeding point based on the hull pose data to obtain a stable video stream after eliminating motion blur. The feeding status label of the aquatic organisms at the feeding point is determined in each video frame of the stable video stream, and then feature extraction is performed on the video frames in the stable video stream based on all feeding status labels to obtain the feeding behavior features of the aquatic organisms in all video frames. Feature decoupling is performed on all feeding behavior features to determine the feeding response index of the aquatic organisms at the feeding point at the current moment. Based on the feeding response index, a stop feeding command is triggered to control the onboard feeder to interrupt the feeding at the feeding point, thereby automatically feeding each feeding point on the cruise path in the aquaculture area.

[0014] Therefore, this application uses the feeding response index to trigger a stop-feeding command to control the shipborne feeder to interrupt feed delivery at feeding points, thereby automatically feeding each feeding point along the cruise path in the aquaculture area. First, determining a stable video stream yields a continuous frame sequence without motion blur after sharpening and enhancing the blurred areas caused by high-frequency vibrations in the original video stream. The determination of a stable video stream eliminates image shifts and motion blur caused by ship translation, rotation, and high-frequency vibrations through affine transformation compensation based on the unmanned vessel's hull pose data, ensuring clear and continuous video frames, which is crucial for subsequent feeding status labeling. Accurate labeling provides a high-quality image foundation. Then, determining feeding behavior features yields a multi-dimensional feature vector that quantifies the intensity of agglomerative feeding behavior in farmed aquatic organisms stimulated by feed. The determination of feeding behavior features, through the fusion of multi-dimensional information such as inverse moment, gray-level difference statistical entropy, and histogram statistics, and after scale calibration to eliminate dimensional differences, achieves a comprehensive characterization of the aquatic organisms' feeding state. This avoids the limitations of single features and solves the problem of inaccurate feeding state identification due to one-sided features in existing technologies. Furthermore, the feeding behavior features, through multi-scale filtering and standardization, can stably capture the feeding state of farmed aquatic organisms at different scales. The feeding response index is adjusted to adapt to different aquatic densities and activity ranges, ensuring the accuracy of the feeding response index calculation. Finally, the feeding response index is determined to be an indicator characterizing the current feeding intensity of aquatic organisms at the feeding point. This index transforms multi-dimensional features into intuitive quantitative indicators, providing a clear threshold for stopping feeding and achieving precise feeding control. The determination of the feeding response index solves the problem in existing technologies where it is difficult to accurately characterize aquatic feeding intensity with a single parameter by transforming multi-dimensional texture feature information and real-time feeding activity trends into intuitive quantitative indicators. This not only reflects the effectiveness of the feature set but also directly... By associating the current feeding activity level of the fish population with the data, a comprehensive assessment of their feeding status is achieved. Furthermore, the determination of the feeding response index provides a basis for the dynamic adjustment of feeding strategies in aquaculture, supporting precise control based on real-time feeding status. This avoids problems such as excessive feed waste, water pollution, or insufficient feeding caused by manual feeding or fixed parameter feeding in existing technologies. It enables the automatic feeding of unmanned vessels to adapt to changes in fish feeding in real time, improving aquaculture efficiency and ecological benefits. In summary, based on the above scheme, intelligent feeding control based on aquatic feeding response can be achieved by combining motion compensation of unmanned vessels with precise identification of feeding status. Attached Figure Description

[0015] Figure 1 This is an exemplary flowchart of an intelligent feeding method for aquaculture based on an unmanned vessel, according to some embodiments of this application. Figure 2 This is an exemplary flowchart illustrating the determination of feeding behavior characteristics according to some embodiments of this application; Figure 3This is a flowchart illustrating the operation of determining the feeding response index according to some embodiments of this application; Figure 4 This is a schematic diagram of the structure of an intelligent feeding system for aquaculture based on an unmanned vessel, according to some embodiments of this application; Figure 5 This is an internal structural diagram of a computer device for implementing an intelligent feeding method for aquaculture based on an unmanned vessel, according to some embodiments of this application. Detailed Implementation

[0016] To better understand the technical solution of this application, the technical solution of this application will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0017] refer to Figure 1 The figure is an exemplary flowchart of an intelligent feeding method for aquaculture based on unmanned vessels, according to some embodiments of this application. The intelligent feeding method for aquaculture based on unmanned vessels mainly includes the following steps: In step 101, the unmanned vessel is controlled to move along the cruise path of the aquaculture area and hover at each feeding point to start feeding.

[0018] In some embodiments, controlling the unmanned vessel to move along the cruise path of the aquaculture area and hover at each feeding point to initiate feed dispensing can be achieved through the following steps: A dynamic path calibration algorithm is used to generate a cruise path for the aquaculture area based on the electronic fence of the target aquaculture area and the preset coordinates of the feeding points. Control the unmanned vessel to move along the cruise path of the aquaculture area; For each feeding point along the cruise path of the aquaculture area, when the unmanned vessel moves along the cruise path of the aquaculture area and arrives at the feeding point, the dual-axis attitude stabilization system of the unmanned vessel is activated to achieve hovering through propeller differential adjustment. Once the unmanned vessel has hovered and stabilized, the onboard feeder is activated by triggering a signal to dispense feed at the preset maximum amount, thereby controlling the unmanned vessel to hover at each feeding point and begin dispensing feed.

[0019] It should be noted that in this application, the aquaculture waterway cruise path is the optimal navigation trajectory for the unmanned vessel to arrive at each feeding point in sequence. This aquaculture waterway cruise path can ensure that the unmanned vessel efficiently traverses all feeding points, avoids waterway obstacles, reduces ineffective navigation, and improves the coverage integrity and path economy of automatic feeding.

[0020] In practice, the dynamic path calibration algorithm can generate a cruise path for the aquaculture area based on the electronic fence of the target aquaculture area and the coordinates of the preset feeding points. This can be achieved in the following way: First, the boundary coordinates of the electronic fence of the target aquaculture area (such as the latitude and longitude of the polygon vertices) and the coordinates of all preset feeding points (such as latitude and longitude) can be obtained. Then, the target aquaculture area is divided into multiple grid cells of the same size (such as 1 meter × 1 meter) and feasible areas and obstacles (such as wooden stakes and shoals) within the electronic fence are marked. Finally, an initial path is constructed with each feeding point as a node using a dynamic path calibration algorithm (such as the improved A* algorithm). The path offset correction coefficient (i.e., path offset correction coefficient = 1 + 0.1 × water flow speed / ship speed) is introduced in combination with the real-time water flow speed (which can be obtained by the ship's onboard current speed sensor) to perform segmented smoothing of the initial path. Finally, a cruise path for the aquaculture area that takes into account both the shortest distance and navigation safety is output.

[0021] In practice, controlling the unmanned vessel to move along the cruise path of the aquaculture area can be achieved in the following way: the waypoint sequence of the cruise path of the aquaculture area can be imported into the shipboard control system. The position deviation calculation module of the shipboard control system can be used to calculate the distance deviation and direction deviation between the current position and the target waypoint and control the unmanned vessel to turn, so as to ensure that it moves along the path.

[0022] In practice, when the unmanned vessel (UAV) moves along the cruise path of the aquaculture area and arrives at the feeding point, the UAV's dual-axis attitude stabilization system is activated to achieve hovering through propeller differential speed adjustment. This can be achieved in the following way: when the straight-line distance between the UAV and the feeding point is less than 5 meters, the UAV's dual-axis attitude stabilization system is activated. This dual-axis attitude stabilization system can apply a speed difference to the left and right propellers and a thrust difference to the front and rear propellers based on the position deviation (such as the straight-line distance between the UAV and the feeding point) through the propeller differential speed adjustment module to counteract the deviation caused by water flow and wind. When the position deviation is less than or equal to 0.5 meters for 3 consecutive seconds, it can be determined that the hovering is stable and a hovering status signal is output.

[0023] In practice, once the unmanned vessel has hovered and stabilized, triggering the onboard feeder to start dispensing feed at the preset maximum feeding amount can be achieved in the following way: After receiving the hovering status signal, the onboard central control module can send a start command to the onboard feeder, controlling the feeder to start the screw conveyor to dispense feed at the preset maximum feeding amount according to the preset conveyor flow rate (e.g., 50 grams per second), thus completing the feeding start control. During the feeding process, the feed delivery amount can be monitored in real time by photoelectric sensors. When the cumulative delivery amount reaches the maximum feeding amount, a deceleration signal is triggered to stop the delivery.

[0024] It should be noted that in this application, the maximum feeding amount is the maximum amount of feed that can be fed to the feeding point at a single time, which is pre-stored in the aquaculture management system. This maximum feeding amount can be set to 50% of the total daily demand of the feeding point. This can not only meet the basic feeding needs of aquatic organisms, but also provide a benchmark for subsequent dynamic adjustments based on feeding status, avoiding the impact of insufficient or excessive feeding on subsequent precise control, and balancing feeding needs and feeding efficiency.

[0025] In step 102, for each feeding point along the cruise route in the aquaculture area, the unmanned vessel collects video streams of the water surface at the feeding point using its onboard camera.

[0026] It should be noted that, in this application, the surface video stream is a continuous sequence of video frames displaying the feeding activities of aquatic organisms in the feeding area. This surface video stream can provide high-quality dynamic image data for subsequent analysis of the feeding status of aquatic organisms, ensuring the accuracy of feeding behavior feature extraction.

[0027] In practice, the video stream of the unmanned vessel at the feeding point can be acquired by the shipborne camera in the following way: The shipborne multispectral anti-shake camera can be automatically adjusted to a 30° tilt angle by rotating the gimbal to ensure that the lens covers the core feeding area with a diameter of 5 meters at the feeding point. The red, green, blue and near-infrared dual-channel video streams are acquired at a frame rate of 25 frames per second. At the same time, the water surface reflection suppression algorithm is enabled to dynamically adjust the exposure parameters (such as exposure time = baseline exposure value × (1 - light spot area ratio × 0.3)) by analyzing the brightness value of the light spot area in the frame in real time to reduce the light spot interference caused by direct sunlight. Invalid frames caused by violent water wave fluctuations are eliminated by comparing the pixel change rate of adjacent frames (such as the change rate threshold can be set to 5%), and finally a continuous and stable water surface video stream is formed.

[0028] In step 103, the hull pose data of the unmanned vessel is acquired synchronously, and then the affine transformation compensation of the water surface video stream of the unmanned vessel at the feeding point is performed based on the hull pose data to obtain a stable video stream after eliminating motion blur.

[0029] It should be noted that, in this application, the hull pose data is real-time dynamic data that displays the three-dimensional coordinates (such as longitude, latitude, and draft) and attitude angles (such as roll angle, pitch angle, and heading angle) of the unmanned vessel when it collects the surface video stream at the feeding point. This hull pose data can provide accurate parameters for affine transformation compensation of the surface video stream, thereby eliminating video motion blur caused by hull swaying and ensuring the stability and accuracy of the video frames used for subsequent feeding status analysis.

[0030] In practice, the synchronous acquisition of the hull pose data of the unmanned vessel can be achieved in the following way: while the unmanned vessel is collecting the water surface video stream, the shipborne navigation satellite system (such as the Beidou satellite navigation system) and the inertial measurement unit (which can be composed of gyroscopes and accelerometers) combined navigation module are activated to synchronously collect (e.g., sampling frequency of 100 Hz) the coordinate information (such as longitude and latitude) and the attitude angles of the hull (such as roll angle, pitch angle, and heading angle) of the unmanned vessel. Then, the acquisition time of each video frame in the water surface video stream is bound with the corresponding pose data through a timestamp alignment algorithm to form a synchronous dataset of "video frame-pose". This synchronous dataset is used as the hull pose data of the unmanned vessel.

[0031] In some embodiments, the following steps can be used to perform affine transformation compensation on the surface video stream of the unmanned vessel at the feeding point based on the hull pose data to obtain a stable video stream after motion blur elimination: Construct an affine transformation matrix based on the hull pose data; Based on the affine transformation matrix, offset correction is performed on each video frame in the surface video stream of the unmanned vessel at the feeding point to obtain a corrected video stream. The corrected video stream is sharpened and enhanced based on the pose change rate of the unmanned vessel in the hull pose data to generate a stable video stream after motion blur is eliminated.

[0032] In specific implementation, the affine transformation matrix can be constructed based on the ship's pose data in the following way: the roll angle, pitch angle, yaw angle, and position deviation of each video frame in the ship's pose data can be extracted. The attitude angles are converted into rotation parameters (e.g., the roll angle corresponds to the rotation matrix around the x-axis, the pitch angle corresponds to the rotation matrix around the y-axis, and the yaw angle corresponds to the rotation matrix around the z-axis). The position deviation is converted into translation parameters. Then, the rotation parameters and translation parameters are fused into a transformation matrix using the affine transformation matrix synthesis formula (e.g., matrix = rotation matrix × scaling factor + translation matrix, where the scaling factor can be set to 1 to keep the proportion unchanged). Each video frame corresponds to a transformation matrix. The set of transformation matrices synchronized with the video frames in the water surface video stream is then output as the affine transformation matrix. The affine transformation matrix is ​​a set of transformation matrices describing the pixel coordinate transformation relationship of the video frames. This affine transformation matrix can provide a mathematical basis for the offset correction of video frames in the water surface video stream, accurately offsetting the image misalignment caused by ship rotation and translation, and laying the foundation for video stabilization processing.

[0033] In specific implementation, the offset correction is performed on each video frame in the surface video stream of the unmanned vessel at the feeding point based on the affine transformation matrix. The corrected video stream can be obtained in the following way: the affine transformation matrix and the surface video stream of the unmanned vessel at the feeding point can be input into the warpAffine function in an existing computer vision processing tool (such as OpenCV) to eliminate the image offset caused by the translation and rotation of the vessel, forming a corrected video stream; wherein, the warpAffine function performs coordinate transformation on each pixel in the video frame based on a single video frame and the corresponding transformation matrix in the surface video stream (such as new coordinates = affine matrix × original coordinates) and calculates the gray value of the transformed pixel through a bilinear interpolation algorithm, thereby controlling the pixel offset of adjacent frames to within 3 pixels, and then arranging all the corrected video frames in time sequence to output a continuous corrected video stream without jumps; wherein, the corrected video stream is a continuous frame sequence formed after eliminating the overall image offset caused by the translation and rotation of the vessel in the original surface video stream.

[0034] In specific implementation, the correction video stream is sharpened and enhanced based on the pose change rate of the unmanned vessel in the hull pose data to generate a stable video stream after motion blur elimination. This can be achieved in the following way: First, for each video frame in the correction video stream, the average value of the time derivative of all attitude angles in the hull pose data corresponding to the video frame can be calculated as the pose change rate of the unmanned vessel. When the pose change rate is greater than a preset change rate threshold (e.g., 0.5 degrees per second), the video frame is determined to be high-frequency vibration. An adaptive Laplacian sharpening algorithm can be used with the video frame as input, and the sharpening kernel strength of the correction video stream can be adjusted according to the pose change rate (e.g., sharpening kernel strength = 0.3 + 0.2 × pose change rate) to enhance the edge region within the video frame. Through the above steps, each video frame in the correction video stream can be sharpened and enhanced. Finally, the continuous frame sequence of all video frames after sharpening and enhancement processing arranged in chronological order is used as the stable video stream after motion blur elimination.

[0035] It should be noted that in this application, the stable video stream is a continuous frame sequence without motion blur obtained by sharpening and enhancing the blurred areas caused by high-frequency vibration in the original video stream. This stable video stream can ensure that the video frames are clear and stable, thereby improving the accuracy of aquatic feeding status identification and providing high-quality data for subsequent feature extraction and feeding response index calculation.

[0036] In step 104, the feeding status label of the farmed aquatic organisms at the feeding point in each video frame of the stable video stream is determined, and then feature extraction is performed on the video frames in the stable video stream based on all the feeding status labels to obtain the feeding behavior features of the farmed aquatic organisms in all video frames.

[0037] In some embodiments, determining the feeding status label of cultured aquatic organisms at feeding points in each video frame of the stable video stream can be achieved using the following steps: The feeding status of farmed aquatic organisms in the stable video stream before the unmanned vessel feeds the aquatic organisms is labeled as low-activity; The feeding status tags of aquatic organisms in the initial video frame of the unmanned vessel feeding in the stable video stream are marked as high-activity, thereby determining the feeding status tags of aquatic organisms at each feeding point in the stable video stream.

[0038] It should be noted that, in this application, the feeding state label is an identifier that marks the feeding activity of aquatic organisms in each video frame of a stable video stream (such as "low activity state" or "high activity state"). This feeding state label can provide a classification benchmark for subsequent extraction of texture features and analysis of feeding state, thereby enabling targeted learning of the feature differences of different activity levels and improving the accuracy of feeding response judgment.

[0039] In specific implementation, the feeding status of aquatic organisms in the stable video stream before the unmanned vessel feeds the aquatic organisms is labeled as "low-activity state" in the video frames. This can be achieved in the following way: using the feeder start signal as the time reference, a stable video stream segment is extracted 2 minutes before the start signal is emitted through the timestamp synchronization mechanism of the onboard system. The feeding status of aquatic organisms in each video frame of this stable video stream segment is labeled as "low-activity state". At the same time, the proportion of fish pixels in each video frame of this stable video stream segment can be calculated by an inter-frame aquatic organism density detection algorithm (such as a target detection algorithm based on background difference). If the aquatic organism density in 10 consecutive video frames is lower than a preset low density threshold (such as 10%), the label is confirmed to be valid. The low-activity state refers to the state in which aquatic organisms are naturally dispersed and have low feeding desire when they are not in contact with feed. This low-activity state serves as a benchmark reference for the feeding state, contrasting with the high-activity state, and can be used to identify the start and intensity changes of feeding behavior, providing an initial reference for subsequent dynamic control.

[0040] In specific implementation, marking the feeding status of aquatic organisms in the initial video frame of the unmanned vessel feeding in the stable video stream as a "high-activity state" can be achieved in the following way: A stable video stream segment within the initial time period (e.g., 0-2 minutes) after the feeder starts can be extracted using the timestamp synchronization mechanism of the onboard system. Based on the aquatic organisms' aggregation and feeding characteristics under feed stimulation, each video frame within this initial time period is marked as "high-activity state." Simultaneously, the percentage of fish pixels in each video frame within this stable video stream segment can be calculated using an inter-frame aquatic organism density detection algorithm (e.g., a target detection algorithm based on background difference). If the aquatic organisms in 10 consecutive video frames are in a "high-activity state,"... If the density is higher than the preset high-density threshold (e.g., 80%), the label is confirmed to be valid. The inter-frame aquaculture density detection algorithm first establishes the background region through a Gaussian mixture model, and then performs a difference operation between the video frame and the background region to extract the foreground fish body region. After removing noise through morphological filtering, the ratio of the total number of pixels in the fish body region to the total number of pixels in the video frame is calculated to obtain the fish body pixel ratio, and the aquaculture density value of each video frame is output. The high-activity state refers to the state in which the corresponding aquatic animals gather and compete for food when stimulated by feed, and have a strong desire to feed. This high-activity state serves as a typical sample of active feeding and can be used to accurately identify the decay process of feeding intensity, ensuring timely adjustment of feeding strategies.

[0041] In some embodiments, reference Figure 2 The figure is an exemplary flowchart illustrating the determination of feeding behavior characteristics according to some embodiments of this application. In this application, feature extraction is performed on video frames in the stable video stream based on all feeding state labels to obtain the feeding behavior characteristics of farmed aquatic organisms in all video frames. This can be achieved by the following steps: In step 1041, multi-scale filtering is performed on the video frames marked with feeding state tags in the stable video stream to obtain a set of effective video frames with texture enhancement. In step 1042, for each video frame in the set of valid video frames, the inverse difference moment of the gray-level co-occurrence matrix in the video frame is extracted; In step 1043, the grayscale difference statistical entropy and histogram of the video frame are determined; In step 1044, multi-dimensional texture features of video frames are generated based on the inverse difference moment, the gray-level difference statistical entropy and the histogram statistics, thereby obtaining the multi-dimensional texture features of each video frame in the effective video frame set. In step 1045, scale calibration is performed on each multi-dimensional texture feature to obtain the feeding behavior features of farmed aquatic products in all video frames.

[0042] In specific implementation, multi-scale filtering of video frames labeled with feeding status tags in the stable video stream to obtain a set of effective video frames with enhanced texture can be achieved in the following way: an adaptive Gabor filter bank can be used to filter each video frame labeled with feeding status tags in the stable video stream. Specifically, five scales (e.g., wavelengths of 3, 5, 7, 9, and 11 pixels) and four directions (e.g., 0°, 45°, 90°, and 135°) are set, and the filtering results in each direction are maximized by calculating the edge density difference between feeding and non-feeding frames at different scales to synthesize texture-enhanced frames. All the filtered video frames are then arranged in chronological order to form a set of effective video frames. The set of effective video frames refers to a high-quality video frame sequence that accurately reflects the feeding status of aquatic animals after being filtered by feeding status tags.

[0043] In practice, the inverse moment of the gray-level co-occurrence matrix in a video frame can be extracted using the following method: A sliding window (e.g., 125 pixels × 125 pixels) can be used to traverse the video frame region by region to calculate the gray-level co-occurrence matrix of the video frame. The inverse moment is then calculated based on the existing formula for inverse moment (e.g., inverse moment = the sum of the products of the squares of the gray-level differences between all pixel pairs and the corresponding matrix elements). The sliding window can start from the top left corner of the video frame and move horizontally sequentially (e.g., with a step size set to half the window side length, i.e., 62 pixels). A fixed distance (e.g., 1 pixel) is statistically analyzed for each window's coverage area. The gray-level co-occurrence matrix of a region is generated by analyzing the distribution of pixel gray-level pairs at fixed distances and directions (e.g., 0°). The inverse moment is an indicator of texture homogeneity in a video frame. The larger the inverse moment value, the more uniform the texture in the video frame. The inverse moment can effectively distinguish between the dispersed (i.e., high inverse moment) and aggregated (i.e., low inverse moment) states of farmed aquatic organisms, thereby assisting in judging feeding activity. The gray-level co-occurrence matrix is ​​a matrix describing the distribution of pixel gray-level pairs at fixed distances and directions in a video frame. It can reflect the spatial relationship of texture in the video frame, thereby quantifying the local texture pattern of aquatic organism aggregation and providing spatial structural feature basis for distinguishing feeding and non-feeding states.

[0044] In practical implementation, determining the gray-level difference statistical entropy and histogram statistics of a video frame can be achieved as follows: First, the gray-level differences between adjacent pixels in the video frame can be calculated and the probability distribution of the gray-level differences can be statistically analyzed. Then, the gray-level difference statistical entropy can be calculated using existing entropy formulas (e.g., entropy = the negative of the sum of the products of the probabilities of all gray-level differences and the natural logarithm of those probabilities). Next, the mean, variance, skewness (i.e., the ratio of the third central moment to the cube of the standard deviation), and kurtosis (i.e., the ratio of the fourth central moment to the fourth power of the standard deviation) of all pixel gray levels in the video frame can be calculated. The vector composed of the mean, variance, skewness, and steepness is used as the histogram statistic. The gray-level difference statistical entropy is an entropy value reflecting the texture complexity in the video frame. This gray-level difference statistical entropy quantifies the dynamic texture features during feeding by the gray-level difference changes caused by the movement of farmed aquatic organisms, thereby enhancing the sensitivity of state recognition. The histogram statistic is a statistical vector reflecting the overall brightness and contrast in the video frame. This histogram statistic can supplement the overall texture distribution features by capturing the gray-level differences between farmed aquatic organisms and the water surface during feeding, thereby improving the completeness of the feature dimensions.

[0045] In specific implementation, the multi-dimensional texture features of video frames generated based on the inverse difference moment, the gray-level difference statistical entropy, and the histogram statistics can be implemented in the following way: the mean, variance, skewness, and steepness of the gray-level co-occurrence matrix inverse difference moment, gray-level difference statistical entropy, and histogram statistics can be arranged in a fixed order (e.g., inverse difference moment → gray-level difference statistical entropy → mean → variance → skewness → steepness) to form a 6-dimensional feature vector, and this 6-dimensional feature vector is used as the multi-dimensional texture features of the video frame; wherein, the multi-dimensional texture features are multi-dimensional feature vectors that comprehensively reflect the local uniformity of texture, texture complexity, and overall brightness and contrast in the video frame. These multi-dimensional texture features can characterize the feeding state of farmed aquatic products from multiple dimensions such as local structure, dynamic changes, and overall distribution, thereby improving the distinguishability of features.

[0046] It should be noted that in this application, the feeding behavior feature is a multi-dimensional feature vector that quantifies the intensity of agglomeration and competitive feeding behavior of farmed aquatic organisms due to feed stimulation. This feeding behavior feature dynamically perceives the feeding activity of fish groups in multiple dimensions, and can adapt to different densities and activity ranges of farmed aquatic organisms, thereby enhancing the feature's adaptability to changes in feeding status. In specific implementation, the feeding behavior features of farmed aquatic organisms in all video frames can be obtained by performing scale calibration on each multi-dimensional texture feature. This can be achieved by using existing standardization algorithms (such as the max-min standardization algorithm) to perform multi-scale calibration on the multi-dimensional texture features of each video frame. That is, the feature values ​​of each dimension of the multi-dimensional texture feature are mapped to the 0-1 interval according to the standardization formula to eliminate the dimensional differences of different features. Then, the calibrated multi-dimensional texture features are used as the feeding behavior features corresponding to each video frame, thereby obtaining the feeding behavior features of farmed aquatic organisms in all video frames.

[0047] In step 105, feature decoupling is performed on all feeding behavior characteristics to determine the feeding response index of the aquatic organisms at the feeding point at the current moment. Based on the feeding response index, a stop feeding command is triggered to control the shipborne feeder to interrupt the feed delivery at the feeding point, thereby automatically feeding each feeding point on the cruise path of the aquaculture area.

[0048] In some embodiments, reference Figure 3 The figure is a flowchart illustrating the operation of determining the feeding response index according to some embodiments of this application. In this application, feature decoupling of all feeding behavior characteristics is performed to determine the feeding response index of aquatic organisms at the feeding point at the current moment. This can be achieved by the following steps: Construct a feature orthogonality matrix based on all feeding behavior characteristics; Determine the proportion of orthogonal features in the feature orthogonality matrix; The proportion of high-activity frames in the stable video stream whose feeding status label is high-activity is determined based on all feeding status labels. The orthogonal feature ratio and the high-activity frame ratio are fused to generate the feeding response index of aquatic organisms at the feeding point at the current moment.

[0049] It should be noted that in this application, the feature orthogonality matrix is ​​a matrix composed of the feature orthogonality between all feeding behavior features. This feature orthogonality matrix can intuitively present the degree of correlation between feeding behavior features, thereby providing structured data for rapid identification of orthogonal feature pairs. Among them, feature orthogonality is an indicator that characterizes the correlation between feeding behavior features. The lower the feature orthogonality value, the less information overlap there is between feeding behavior features, which can ensure that the extracted features can reflect the feeding state from different dimensions and reduce information redundancy.

[0050] In practice, the orthogonality matrix of features based on all feeding behavior features can be constructed in the following way: any two feeding behavior features can be combined into a feature pair, thus obtaining multiple feature pairs. For each feature pair, the cosine similarity algorithm (i.e., cosine similarity = dot product of two feature vectors ÷ product of the magnitudes of two feature vectors) is used to calculate the cosine similarity between the feeding behavior features as the orthogonality of the feature pair. Through the above steps, the orthogonality of all feature pairs can be obtained. Then, the orthogonality of all feature pairs is arranged according to the index of the feeding behavior features to construct the orthogonality matrix (i.e., the diagonal elements are set to 1, and the rest are the orthogonality of the corresponding feature pairs).

[0051] In specific implementation, the proportion of orthogonal features in the orthogonality matrix can be determined as follows: feature pairs in the orthogonality matrix with orthogonality below a set orthogonality judgment threshold (e.g., 0.3) can be considered as orthogonal feature pairs. The number of orthogonal feature pairs in the orthogonality matrix is ​​counted, and the ratio of the number of orthogonal feature pairs to the total number of feature pairs is taken as the orthogonal feature proportion. The orthogonal feature proportion is the proportion of feature pairs in the orthogonality matrix that satisfy the orthogonality condition to the total number of feature pairs. This orthogonal feature proportion reflects the independence among all feeding behavior features and the information effectiveness of the feature set. The higher the value of the orthogonal feature proportion, the stronger the feature discrimination among all feeding behavior features, which can provide a high-quality basis for judging the feeding state.

[0052] In specific implementation, determining the proportion of highly active frames in the stable video stream with the "highly active" feeding status label based on all feeding status labels can be achieved in the following way: A sliding window mechanism (e.g., a window of 125 frames) can be used to count the number of video frames with the "highly active" feeding status label in the stable video stream under all windows and calculate the average of their proportions to the total number of video frames as the proportion of highly active frames (i.e., proportion of highly active frames = number of video frames with the "highly active" label ÷ total number of video frames in the stable video stream). The proportion of highly active frames is the ratio of the number of video frames in the stable video stream where aquatic organisms are in a highly active state to the total number of video frames. This proportion directly reflects the overall trend of current feeding activity, quantifies the current feeding activity of aquatic organisms, provides real-time reference for dynamically adjusting feeding strategies, and avoids delayed judgment.

[0053] It should be noted that in this application, the feeding response index is an indicator characterizing the current feeding intensity of aquatic organisms at the feeding point. This feeding response index transforms multi-dimensional features into intuitive quantitative indicators, providing a clear threshold for stopping feeding decisions and achieving precise feeding control. Specifically, the feeding response index of aquatic organisms at the feeding point at the current moment can be generated by feature fusion of the orthogonal feature ratio and the high-activity frame ratio using the following method: First, the orthogonal feature ratio and the high-activity frame ratio can be normalized (e.g., maximum-minimum normalization) to ensure that both are within the 0-1 range; then… Then, the orthogonal feature proportion and the high-activity frame proportion can be fused using the Dempster-Schaffer evidence theory to obtain the feeding response index of aquatic organisms at the feeding point at the current moment. That is, the orthogonal feature proportion and the high-activity frame proportion are used as two independent evidence sources to construct basic probability allocation functions (such as the orthogonal feature proportion corresponding to feature reliability evidence, and the high-activity frame proportion corresponding to real-time state evidence). The comprehensive confidence level is obtained by fusing the evidence through evidence synthesis rules (such as the combination formula being the joint probability of the two pieces of evidence divided by the normalization coefficient). The comprehensive confidence level is then mapped to the 0-1 interval as the feeding response index.

[0054] In some embodiments, the automatic feeding of each feeding point along the cruise path of the aquaculture area can be achieved by triggering a stop feeding command based on the feeding response index to control the shipborne feeder to interrupt feed delivery at the feeding point, thereby enabling automatic feeding: When the feeding response index is lower than a preset response threshold, a stop feeding command is triggered; Based on the stop feeding command, the shipborne feeder is controlled to interrupt the feed delivery at the feeding point, thus completing the control of the current feeding point; After feeding at the current feeding point is completed, the unmanned vessel is controlled to proceed to the next feeding point along the cruise route of the aquaculture area, repeating the above steps to automatically feed each feeding point along the cruise route of the aquaculture area.

[0055] It should be noted that in this application, the stop feeding command is a control signal generated by the shipboard control system to trigger the feeder to interrupt the feed delivery at the current feeding point. This stop feeding command can be used to automatically terminate feeding by monitoring the feeding activity of aquatic animals in real time, thereby avoiding waste and water pollution caused by excessive feed delivery and achieving precise feeding control.

[0056] In specific implementation, when the feeding response index is lower than a preset response threshold, triggering a stop-feeding command can be achieved in the following way: a response threshold for the feeding response index (e.g., 0.4) can be preset through a remote platform based on the characteristics of the cultured aquatic species. When the shipborne control system monitors in real time that the feeding response index is lower than the response threshold for 5 consecutive seconds, it determines that the feeding activity of the cultured aquatic species is insufficient and automatically generates a stop-feeding command. The response threshold is a preset critical value of the feeding response index, which can be flexibly adjusted according to the feeding characteristics of different cultured species. This response threshold can be used to determine whether the feeding activity of the cultured aquatic species is insufficient as a decision basis for triggering a stop-feeding command, thereby providing a quantitative standard for terminating feeding and improving the adaptability and scientific nature of the feeding strategy.

[0057] In specific implementation, the feed delivery at the interrupted feeding point is controlled by the shipborne feeder based on the stop feeding command. The control of the current feeding point can be achieved in the following way: After receiving the stop feeding command, the shipborne feeder controller immediately triggers the electromagnetic gate valve deceleration program to control the screw conveyor speed to linearly decrease from the rated value to 0, and then cuts off the motor power. At the same time, feed residue can be detected by the photoelectric sensor installed at the discharge port. If no feed passes through for 3 consecutive seconds, it is determined that the feeding interruption is completed, and a "current feeding point ended" signal is fed back to the shipborne system and the total amount of feed delivered this time is recorded.

[0058] In another aspect, in some embodiments, this application provides an intelligent feeding system for aquaculture based on an unmanned vessel, see reference. Figure 4 The figure is a schematic diagram of the structure of an intelligent feeding system for aquaculture based on an unmanned vessel, according to some embodiments of this application. The intelligent feeding system 400 for aquaculture based on an unmanned vessel includes: a data acquisition module 401, a processing module 402, and an execution module 403, which are described below: The data acquisition module 401 in this application is mainly used to control the unmanned vessel to move along the cruise path of the aquaculture area and hover at each feeding point to start the feed delivery. It should be noted that the acquisition module 401 in this application is also used to acquire video streams of the unmanned vessel on the water surface at each feeding point along the patrol route of the aquaculture area through the shipborne camera. Processing module 402, in this application, is mainly used to synchronously acquire the hull pose data of the unmanned vessel, and then perform affine transformation compensation on the water surface video stream of the unmanned vessel at the feeding point based on the hull pose data to obtain a stable video stream after eliminating motion blur. It should be noted that the processing module 402 in this application is also used to determine the feeding status label of the aquatic organisms at the feeding point in each video frame of the stable video stream, and then extract features from the video frames in the stable video stream based on all the feeding status labels to obtain the feeding behavior features of the aquatic organisms in all video frames. The execution module 403 in this application is mainly used to decouple all feeding behavior characteristics, thereby determining the feeding response index of the aquatic organisms at the feeding point at the current moment, triggering a stop feeding command based on the feeding response index to control the shipborne feeder to interrupt the feed delivery at the feeding point, and then automatically feeding each feeding point on the cruise path of the aquaculture water area.

[0059] The modules in the aforementioned intelligent feeding system for aquaculture based on unmanned vessels can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the computer device's memory as software, so that the processor can call and execute the corresponding operations of each module.

[0060] In another embodiment, this application provides a computer device, which may be a server, and its internal structure diagram may be as follows. Figure 5 As shown, the computer device includes a processor, memory, and a network interface connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The database stores data on intelligent feeding methods for aquaculture based on unmanned vessels. The network interface communicates with external terminals via a network connection. When the computer program is executed by the processor, it implements an intelligent feeding method for aquaculture based on unmanned vessels.

[0061] Those skilled in the art will understand that Figure 5 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0062] In one embodiment, a computer device is also provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above embodiment of the intelligent feeding method for aquaculture based on unmanned vessels.

[0063] In one embodiment, a computer-readable storage medium is provided storing a computer program that, when executed by a processor, implements the steps described in the embodiment of the intelligent feeding method for aquaculture based on unmanned vessels.

[0064] In one embodiment, a computer program product or computer program is provided, comprising computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the steps described in the embodiment of the intelligent feeding method for aquaculture based on an unmanned vessel.

[0065] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the methods described above. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, or optical storage, etc. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM), etc.

[0066] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0067] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. A smart feeding method for aquaculture based on unmanned vessels, wherein, Automatic feeding control based on the feeding status of farmed aquatic organisms, characterized by the following steps: Control the unmanned vessel to move along the cruise path of the aquaculture area and hover at each feeding point to start feeding; For each feeding point along the patrol route of the aquaculture area, the unmanned vessel collects video streams of the water surface at the feeding point using shipborne cameras. The hull pose data of the unmanned vessel is acquired synchronously, and then the affine transformation compensation of the water surface video stream of the unmanned vessel at the feeding point is performed based on the hull pose data to obtain a stable video stream after eliminating motion blur. The feeding status label of the farmed aquatic organisms at the feeding point is determined in each video frame of the stable video stream. Then, based on all the feeding status labels, feature extraction is performed on the video frames in the stable video stream to obtain the feeding behavior features of the farmed aquatic organisms in all video frames. All feeding behavior characteristics are decoupled to determine the feeding response index of aquatic organisms at the feeding point at the current moment. Based on the feeding response index, a stop feeding command is triggered to control the shipborne feeder to interrupt the feed delivery at the feeding point, thereby automatically feeding each feeding point along the cruise path of the aquaculture area.

2. The method as described in claim 1, characterized in that, Controlling the unmanned vessel to move along the cruise path of the aquaculture area and hover at each feeding point to initiate feed dispensing specifically includes: A dynamic path calibration algorithm is used to generate a cruise path for the aquaculture area based on the electronic fence of the target aquaculture area and the preset coordinates of the feeding points. Control the unmanned vessel to move along the cruise path of the aquaculture area; For each feeding point along the cruise path of the aquaculture area, when the unmanned vessel moves along the cruise path of the aquaculture area and arrives at the feeding point, the dual-axis attitude stabilization system of the unmanned vessel is activated to achieve hovering through propeller differential adjustment. Once the unmanned vessel has hovered and stabilized, it triggers the onboard feed dispenser to start dispensing feed at the preset maximum amount, thereby controlling the unmanned vessel to hover at each feeding point and start dispensing feed.

3. The method as described in claim 1, characterized in that, Based on the ship's pose data, affine transformation compensation is performed on the surface video stream of the unmanned vessel at the feeding point to obtain a stable video stream after motion blur elimination, specifically including: Construct an affine transformation matrix based on the hull pose data; Based on the affine transformation matrix, offset correction is performed on each video frame in the surface video stream of the unmanned vessel at the feeding point to obtain a corrected video stream. The corrected video stream is sharpened and enhanced based on the pose change rate of the unmanned vessel in the hull pose data to generate a stable video stream after motion blur is eliminated.

4. The method as described in claim 1, characterized in that, Determining the feeding status label of aquatic organisms at each feeding point in the stable video stream specifically includes: The feeding status of farmed aquatic organisms in the stable video stream before the unmanned vessel feeds the aquatic organisms is labeled as low-activity. The feeding status tags of aquatic organisms in the initial video frame of the unmanned vessel feeding in the stable video stream are marked as high-activity, thereby determining the feeding status tags of aquatic organisms at each feeding point in the stable video stream.

5. The method as described in claim 1, characterized in that, Based on all feeding state labels, feature extraction is performed on video frames in the stable video stream to obtain the feeding behavior features of farmed aquatic animals in all video frames, specifically including: Multi-scale filtering is performed on video frames marked with feeding status tags in the stable video stream to obtain a set of effective video frames with enhanced texture. For each video frame in the set of valid video frames, extract the inverse difference moment of the gray-level co-occurrence matrix in the video frame; Determine the grayscale difference statistical entropy and histogram of video frames; Based on the inverse difference moment, the gray-level difference statistical entropy and the histogram, multi-dimensional texture features of video frames are generated, and then multi-dimensional texture features of each video frame in the effective video frame set are obtained. Scale calibration was performed on various multi-dimensional texture features to obtain the feeding behavior features of farmed aquatic products in all video frames.

6. The method as described in claim 1, characterized in that, Feature decoupling is performed on all feeding behavior characteristics to determine the feeding response index of aquatic organisms at the feeding point at the current moment. Specifically, this includes: Construct a feature orthogonality matrix based on all feeding behavior characteristics; Determine the proportion of orthogonal features in the orthogonality matrix; The proportion of highly active frames in the stable video stream whose feeding state label is "highly active" is determined based on all feeding state labels. The orthogonal feature ratio and the high-activity frame ratio are fused to generate the feeding response index of aquatic organisms at the feeding point at the current moment.

7. The method as described in claim 1, characterized in that, Based on the feeding response index, a stop-feeding command is triggered to control the shipborne feeder to interrupt feed delivery at feeding points, thereby automatically feeding each feeding point along the cruise path of the aquaculture area. Specifically, this includes: When the feeding response index is lower than a preset response threshold, a stop feeding command is triggered; Based on the stop feeding command, the shipborne feeder is controlled to interrupt the feed delivery at the feeding point, thus completing the control of the current feeding point; After feeding at the current feeding point is completed, the unmanned vessel is controlled to proceed to the next feeding point along the cruise route of the aquaculture area, repeating the above steps to automatically feed each feeding point along the cruise route of the aquaculture area.

8. An intelligent feeding system for aquaculture based on an unmanned vessel, characterized in that, The system includes: The data acquisition module is used to control the unmanned vessel to move along the cruise path of the aquaculture area and hover at each feeding point to start feeding. The acquisition module is used to collect video streams of the unmanned vessel on the water surface at each feeding point along the cruise route of the aquaculture area using a ship-mounted camera. The processing module is used to synchronously acquire the hull pose data of the unmanned vessel, and then perform affine transformation compensation on the water surface video stream of the unmanned vessel at the feeding point based on the hull pose data to obtain a stable video stream after eliminating motion blur. The processing module is used to determine the feeding status label of the farmed aquatic organisms at the feeding point in each video frame of the stable video stream, and then extract features from the video frames in the stable video stream based on all the feeding status labels to obtain the feeding behavior features of the farmed aquatic organisms in all video frames. The execution module is used to decouple all feeding behavior characteristics, thereby determining the feeding response index of aquatic organisms at the feeding point at the current moment. Based on the feeding response index, a stop feeding command is triggered to control the shipborne feeder to interrupt the feed delivery at the feeding point, thereby automatically feeding each feeding point along the cruise path of the aquaculture area.

Images