Intelligent fishing method, device, system and storage medium

Abstract

The application provides a smart fishing method, device, system and storage medium, which comprises the following steps: acquiring a target image sequence of an underwater environment in a water area containing fish to be detected; sequentially performing feature extraction and feature classification processing based on the target image sequence to determine a target recognition result of a target fish; determining carrier pose data of a smart fishing system according to target pixel coordinates corresponding to the target recognition result and the target image sequence, and acquiring a historical position sequence and a target motion feature of the target fish, and inertial measurement unit data of the smart fishing system; performing filtering processing based on the carrier pose data, the inertial measurement unit data, the historical position sequence and the target motion feature to determine a target positioning result and a target prediction trajectory result of the target fish; and controlling a harpoon launching device to release a harpoon for fishing operation based on the target recognition result, the target positioning result and the target prediction trajectory result, so as to improve the accuracy of fishing.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of fishing, and in particular to an intelligent fishing method, device, system and storage medium. BACKGROUND

[0002] Underwater fishing operations have important application value in the fields of aquaculture, water ecological protection and water environment governance. Traditional fishing methods mainly rely on manual diving fishing or trawl operations, which have problems such as low efficiency, poor selectivity, and great disturbance to the water ecology: manual fishing is limited in range and high in cost due to environmental restrictions such as water depth and water flow; although trawl operations are more efficient, they can cause damage to adult fish and mis-capture of juvenile fish, leading to an increase in the loss rate of cultivation, and serious damage to the ecological community on the water bottom.

[0003] With the development of automation technology, existing underwater fishing equipment has begun to transform towards intelligence, but in complex dynamic underwater environments, the accuracy of fishing cannot be guaranteed. SUMMARY

[0004] The present application provides an intelligent fishing method, device, system and storage medium, which improves the accuracy of fishing.

[0005] In a first aspect, an intelligent fishing method is provided, applied to an intelligent fishing system, the intelligent fishing system at least comprising a target recognition module, a target positioning module and a harpoon launching device, and the method comprising:

[0006] obtaining a target image sequence of an underwater environment in a water area containing fish to be detected;

[0007] controlling the target recognition module to sequentially perform feature extraction and feature classification processing based on the target image sequence to determine a target recognition result of a target fish;

[0008] determining carrier pose data of the intelligent fishing system according to target pixel coordinates corresponding to the target recognition result and the target image sequence, and obtaining a historical position sequence and a target motion feature of the target fish, and inertial measurement unit data of the intelligent fishing system;

[0009] controlling the target positioning module to perform filtering processing based on the carrier pose data, the inertial measurement unit data, the historical position sequence and the target motion feature to determine a target positioning result and a target prediction trajectory result of the target fish;

[0010] based on the target recognition result, the target positioning result and the target prediction trajectory result, controlling the harpoon launching device to release a harpoon for fishing operations.

[0011] In a second aspect, an intelligent fishing device is provided, which is applied to an intelligent fishing system, and the intelligent fishing system at least comprises a target recognition module, a target positioning module and a harpoon launching device, and the device comprises:

[0012] an acquisition module, configured to acquire a target image sequence of an underwater environment in a water area containing fish to be detected;

[0013] a first processing module, configured to control the target recognition module to sequentially perform feature extraction and feature classification processing based on the target image sequence, and determine a target recognition result of a target fish;

[0014] a second processing module, configured to determine carrier pose data of the intelligent fishing system according to target pixel coordinates corresponding to the target recognition result and the target image sequence, and acquire a historical position sequence and a target motion feature of the target fish, and inertial measurement unit data of the intelligent fishing system;

[0015] a third processing module, configured to control the target positioning module to perform filtering processing based on the carrier pose data, the inertial measurement unit data, the historical position sequence and the target motion feature, and determine a target positioning result and a target prediction trajectory result of the target fish;

[0016] a control module, configured to control the harpoon launching device to release a harpoon for fishing operation based on the target recognition result, the target positioning result and the target prediction trajectory result.

[0017] In a third aspect, an intelligent fishing system is provided, and the intelligent fishing system comprises:

[0018] a harpoon launching device, a steering engine and a buckle device; wherein

[0019] the harpoon launching device comprises a harpoon and an elastic structure, one end of the elastic structure is connected with the harpoon, and the elastic structure releases elastic potential energy to enable the harpoon launching device to release the harpoon;

[0020] an output shaft of the steering engine is connected with a rotating shaft of the buckle device, and the steering engine is configured to control the buckle device to rotate;

[0021] one end of the buckle device is in abutment with the other end of the elastic structure, and the buckle device controls the elastic structure to release or store elastic potential energy through rotation of the rotating shaft.

[0022] In a fourth aspect, an intelligent fishing system is provided, and the intelligent fishing system comprises:

[0023] a memory, configured to store executable program codes;

[0024] a processor configured to call and run the executable program code from the memory, so that the fishing system executes the intelligent fishing method according to any one of the above.

[0025] In a fifth aspect, a computer readable storage medium is provided, which stores a computer program. When the computer program is executed, the intelligent fishing method according to any one of the above is implemented.

[0026] The technical scheme provided by some embodiments of the present application has at least the following beneficial effects: the intelligent fishing method provided by the present application acquires a target image sequence of an underwater environment in a water area containing fish to be detected; performs feature extraction and feature classification processing based on the target image sequence in sequence to determine a target recognition result of a target fish; determines carrier pose data of an intelligent fishing system according to target pixel coordinates corresponding to the target recognition result and the target image sequence, and acquires a historical position sequence and a target motion feature of the target fish, and inertial measurement unit data of the intelligent fishing system; performs filtering processing based on the carrier pose data, the inertial measurement unit data, the historical position sequence, and the target motion feature to determine a target positioning result and a target prediction trajectory result of the target fish; and controls a harpoon launching device to release a harpoon for fishing operation based on the target recognition result, the target positioning result, and the target prediction trajectory result, so as to improve the accuracy of fishing. BRIEF DESCRIPTION OF DRAWINGS

[0027] In order to more clearly illustrate the technical schemes in the embodiments of the present application or the prior art, the drawings needed in the embodiments or prior art description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0028] Figure 1 is a first flowchart of the intelligent fishing method provided by the embodiments of the present application.

[0029] Figure 2 is a second flowchart of the intelligent fishing method provided by the embodiments of the present application.

[0030] Figure 3 is a third flowchart of the intelligent fishing method provided by the embodiments of the present application.

[0031] Figure 4 is a fourth flowchart of the intelligent fishing method provided by the embodiments of the present application.

[0032] Figure 5 is a fifth flowchart of the intelligent fishing method provided by the embodiments of the present application.

[0033] Figure 6 is a sixth flow diagram of the intelligent fishing method provided by the embodiments of the present application.

[0034] Figure 7 is a seventh flow diagram of the intelligent fishing method provided by the embodiments of the present application.

[0035] Figure 8 is a structural diagram of the intelligent fishing device provided by the embodiments of the present application.

[0036] Figure 9 is a first structural diagram of the intelligent fishing system provided by the embodiments of the present application.

[0037] Figure 10 is a second structural diagram of the intelligent fishing system provided by the embodiments of the present application.

[0038] Figure 11 is a structural diagram of the fish spear launching device provided by the embodiments of the present application.

[0039] Figure 12 is a third structural diagram of the intelligent fishing system provided by the embodiments of the present application. DETAILED DESCRIPTION

[0040] In order to make the features and advantages of the present application more obvious and easy to understand, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, but not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0041] The following description refers to the accompanying drawings. Unless otherwise indicated, same or similar elements in different drawings are denoted by same or similar reference numerals. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of the present application as detailed in the appended claims.

[0042] Hereinafter, the terms "first" and "second" are used only for descriptive purposes and should not be construed as implying or suggesting relative importance or implicitly indicating the number of the indicated technical features. Therefore, the features defined with "first" and "second" can explicitly or implicitly include one or more of the features.

[0043] In order to improve the accuracy of intelligent fishing, the embodiments of the present application provide an intelligent fishing method, which is described in detail below. It should be noted that the description order of the following embodiments is not limited to the preferred order of the embodiments.

[0044] Referring to Figure 1 , Figure 1 is a first flowchart of an intelligent fishing method provided by an embodiment of the present application. The execution subject of the intelligent fishing method can be an intelligent fishing system, which at least includes a target identification module, a target positioning module, and a fish spear launching device. The specific flow of the intelligent fishing method can be as follows:

[0045] S101, obtaining a target image sequence of an underwater environment in a to-be-detected water area containing fish.

[0046] In this embodiment, the intelligent fishing system is a system capable of operating underwater and performing fishing. The intelligent fishing system can be used as a kind of underwater carrier, such as an underwater unmanned aerial vehicle. The intelligent fishing system can perform fishing operations on the to-be-detected water area, which can be any water area such as a river, a lake, or a sea containing fish and other organisms. The target image sequence is a sequence of continuous multiple images obtained in the to-be-detected water area.

[0047] In some embodiments, the step of obtaining a target image sequence of an underwater environment in a to-be-detected water area containing fish can include: obtaining an initial image sequence of the underwater environment in the to-be-detected water area; and performing despeckling processing on the initial image sequence based on an optical compensation algorithm sub-module to eliminate speckle noise in the initial image sequence caused by turbid medium, to obtain the target image sequence.

[0048] Specifically, the intelligent fishing system is provided with an underwater shooting device. The initial image sequence obtained by the underwater shooting device is an image sequence shot in a complex dynamic underwater environment containing turbid water, variable lighting, and other challenges, which is used for subsequent identification of target fish. Based on the optical compensation algorithm sub-module, the initial image sequence is despeckled to eliminate the speckle noise interference in the initial image sequence caused by turbid medium, restore the original optical characteristics of the underwater target, and obtain a relatively clear target image sequence.

[0049] S102, controlling the target identification module to sequentially perform feature extraction and feature classification processing based on the target image sequence to determine a target identification result of the target fish.

[0050] In some embodiments, the target recognition module further comprises a scale perception backbone network submodule, a dual-stream spatio-temporal attention network submodule, a multi-scale feature blunting network submodule, and a multi-level serial classification submodule, wherein the dual-stream spatio-temporal attention network submodule comprises a spatial feature stream attention module, a temporal feature stream attention module, and a dual-stream feature integration module. It should be noted that this embodiment builds a complete recognition process by fusing multiple technical means, covering optical compensation, feature extraction network, multi-level classification, and adaptive mechanism, to solve the recognition problem in complex environments. It can help underwater fishery resource monitoring, ecological research, and other scenarios, accurately identify target fish, improve the reliability and practicality of underwater visual recognition in complex environments, and promote the development of underwater intelligent monitoring technology.

[0051] To accurately distinguish specific target fish and non-target organisms and background in complex dynamic underwater environments containing turbid water bodies, variable lighting, and other challenges, and automatically classify and confirm according to pre-set shape, texture, motion pattern, and other characteristics, in an optional embodiment, the control target recognition module sequentially performs feature extraction and feature classification processing based on the target image sequence, and the step of determining the target recognition result of the target fish can include:

[0052] The scale perception backbone network submodule extracts features from the target image sequence in spatial and temporal dimensions, respectively, to obtain spatio-temporal fusion feature maps of the fish; the spatial feature stream attention module is controlled to perform weighted processing on the spatio-temporal fusion feature maps in the spatial dimension to obtain target morphological features of the fish; the temporal feature stream attention module is controlled to obtain multiple motion features of the spatio-temporal fusion feature maps in the temporal dimension to obtain target motion trajectory features of the fish; the dual-stream feature integration module is used for cross-dimensional feature integration of the target morphological features and the target motion trajectory features to obtain an initial feature vector of the fish; the multi-scale feature blunting network submodule is controlled to sequentially perform multi-scale feature extraction and feature blunting processing on the initial feature vector to compress background feature interference to obtain a target feature vector; and the multi-level serial classification submodule is used for multi-level feature classification of the target feature vector to determine the target recognition result of the target fish.

[0053] It should be noted that the scale perception backbone network submodule is a lightweight scale perception backbone network submodule. The scale perception backbone network submodule efficiently extracts static morphological features such as scales and body contours of the fish through a lightweight network structure with reduced calculation parameters, and simultaneously fuses image frame information at different times to generate spatio-temporal fusion feature maps containing changes in the time dimension, which not only preserves the detailed morphology of the target but also reflects its dynamic change trend.

[0054] Specifically, the scale pattern perception backbone network submodule extracts features of the target image sequence processed by the underwater optical compensation algorithm through a convolution layer. The convolution kernel slides on the target image sequence to extract spatial information of the fish target, such as the texture pattern of the fish scales and the contour of the fish body.

[0055] Specifically, the scale pattern perception backbone network submodule combines structures such as recurrent neural networks or temporal convolution networks to capture the state changes of the fish at different times in the target image sequence, such as changes in the swimming posture of the fish, and fuses the spatial features at different times in the time dimension to obtain spatiotemporal fusion feature maps of the fish containing both spatial information and temporal information.

[0056] It should be noted that the dual-flow spatiotemporal attention network submodule is divided into two branches and processed in parallel: one branch focuses on morphological features and extracts static features such as the length-width ratio of the fish body and the scale texture entropy value; the other branch captures motion features, including dynamic motion trajectory features such as tail swing frequency and turning angular velocity.

[0057] Specifically, the spatial feature flow attention module focuses on spatial details within the static image frame and performs weighted processing on spatial dimension information such as scale texture and fish body contour in the spatiotemporal fusion feature map through convolution, pooling, and other operations to suppress background interference and obtain target morphological features that highlight the texture and shape pattern of the fish. For example, the spatial feature flow attention module gives higher response to the scale texture region to assist in calculating the scale texture entropy value; the spatial feature flow attention module frames the fish body contour region and calculates the length-width ratio of the fish body in combination with geometric transformation.

[0058] Specifically, the time feature flow attention module extracts motion change information between adjacent frames or consecutive frames in the target image sequence, such as local displacement of the fish body during tail swing and contour angle difference during turning, to obtain target motion trajectory features of the fish based on multiple motion features in the motion change information. For example, the time feature flow attention module tracks the trend of feature changes, counts the number of tail swings per unit time to determine the tail swing frequency, and calculates the direction angle difference of the fish body between adjacent frames to determine the turning angular velocity.

[0059] Specifically, the dual-flow feature integration module integrates the target morphological features and the target motion trajectory features of the fish using feature splicing or attention weighted fusion to cross-dimensionally fuse the static morphological features and the dynamic motion trajectory features, so as to obtain an initial feature vector of the fish containing both static morphological features and dynamic motion trajectory features.

[0060] It should be noted that the multi-scale feature blunting network submodule filters out background interference information such as water flow and impurities through multi-scale filtering and feature screening to output a high-dimensional feature vector containing key features of the fish.

[0061] Specifically, the multi-scale feature extraction network submodule performs multi-scale feature extraction on the initial feature vector. A plurality of convolution kernels with different receptive fields are arranged in the multi-scale feature extraction network submodule. For example, a small-scale convolution kernel captures fine texture, and a large-scale convolution kernel captures overall contour, so as to perform multi-scale convolution operation on the initial feature vector. Convolution of different scales can extract target and background features with different sizes and ranges from the initial feature vector. For example, a small-scale convolution focuses on local scale details of a fish body, and a large-scale convolution focuses on the overall distribution of the fish body in the scene and the large-scale correlation with the background.

[0062] The multi-scale feature extraction network submodule performs feature passivation processing on the initial feature vector. Channel attention and spatial attention mechanisms are introduced to perform weighted processing on the initial feature vector after multi-scale feature extraction. The channel attention mechanism focuses on the importance of different feature channels. For underwater target recognition, the channel attention mechanism can strengthen feature channels related to fish morphology and movement, such as scale texture and tail movement channels, and suppress background-related feature channels, such as pure water background and meaningless environmental noise channels. The spatial attention mechanism strengthens the regions where key features are located in the spatial dimension, such as fish body contour and motion trajectory coverage regions, and weakens background regions, such as blank water and fixed rock background regions, thereby compressing background interference and reducing background interference.

[0063] The initial feature vector after multi-scale feature extraction, background suppression, and feature purification is subjected to global pooling operation, and the initial feature vector is subjected to dimension adjustment and feature integration to obtain a high-dimensional feature vector, i.e., a target feature vector.

[0064] It should be noted that the target feature vector is input into the multi-stage serial classification submodule for multi-stage feature classification to determine the target recognition result of the target fish, so that the recognition accuracy can be ensured in a complex underwater environment, and the efficiency can be improved through step-by-step filtering, so that the target recognition result can be directly used for accurate classification in an actual scene.

[0065] Optionally, the multi-stage serial classification submodule can include a four-stage serial classifier chain. The first-stage classifier chain performs target / background binary classification. The second-stage classifier chain implements fish / non-fish biological coarse classification through a lightweight classification network obtained by transfer learning. The third-stage classifier chain completes species fine-grained recognition using a plurality of edge detection operators in combination with a support vector machine. The fourth-stage classifier chain evaluates the target priority based on a body size coefficient and a motion aggressiveness index.

[0066] Specifically, the first-level classifier chain uses simple and efficient binary classifiers to determine whether the target feature vector contains biological feature signals. Based on whether there are continuous motion trajectories, regular contours such as the streamlined contour of a fish body, dynamic texture changes such as texture displacement caused by tail swinging, and other features, the target feature vector is distinguished from pure background features such as stationary rocks, uniform water body turbidity texture, and random motion of bubbles, thereby marking the target or background. Only features with target labels enter the second-level classifier chain, directly filtering 70-80% of pure background interference, reducing subsequent computational load, and improving recognition accuracy.

[0067] For example, if the target feature vector detects a closed contour similar to a fish body for 5 consecutive frames, and a periodic tail swinging motion feature of 1-2 times per second, the feature is determined as a target. If the target feature vector detects irregular contours and no periodic motion such as the stationary features of a rock, or random flickering spots such as background noise caused by sunlight refraction, the feature is determined as background.

[0068] Specifically, the second-level classifier chain uses a lightweight classification network trained by transfer learning to focus on the core feature differences of living organisms. Among them, the presence of scale texture, i.e., high-entropy texture channels in the target feature vector, fin structure, i.e., multiple small-area protruding contours displayed in edge detection, and spindle or laterally flattened body shape, i.e., fish body length-width ratio between 2-5, are fish features. Non-fish biological features include no scales, i.e., low texture entropy, hard shells such as crab shells, i.e., high hardness edge features displayed in the target feature vector, and multi-legged / tentacles such as octopus tentacles, i.e., features displaying multi-branch motion trajectories. Marking fish or non-fish organisms, non-fish organisms are temporarily stored in the second-level classifier chain after being marked, and only features with fish labels enter the third-level classifier chain.

[0069] For example, if the scale texture entropy value in the target feature vector is greater than 0.8, the fish body length-width ratio is greater than 3.2, and there are pectoral fin / tail fin edge features, the feature is determined as a fish. If the target feature vector has no scale, i.e., texture entropy is less than 0.3, multi-branch motion trajectories, and a circular body shape, such as a length-width ratio of less than 1.1, the feature is determined as a non-fish organism such as an octopus.

[0070] Specifically, the third-level classifier chain combines multiple edge detection operators and support vector machines to identify species at a fine granularity. The edge detection operator extracts edge features of key structures such as fish dorsal fin shape and tail fin type, such as the triangular shape of a shark's dorsal fin and the dorsal fin shape of a tuna. The tail fin type includes sickle-shaped, round tail, and bifurcated tail. The support vector machine compares the extracted edge features with a feature library of known species through a trained classification hyperplane to achieve fine classification, thereby achieving accurate identification from fish to specific species.

[0071] For example, if the edge detection shows that the fish has a triangular dorsal fin, a serrated tail fin, and a streamlined body shape, the support vector machine matches the fish to be a shark; if the edge detection shows that the fish has a sickle-shaped dorsal fin, a deep forked tail fin, and silver scales, the support vector machine matches the fish to be a tuna.

[0072] Specifically, the fourth-level classifier chain evaluates the target priority according to a body shape coefficient and a motion aggressiveness index, wherein the body shape coefficient corresponds to the product of the body length, the body width, and the body weight coefficient, and can reflect the size of the target body shape; the motion aggressiveness index corresponds to the tail swing frequency and the turning angle speed, and the higher the frequency and the more rapid the turning, the higher the aggressiveness or activity. Optionally, the fourth-level classifier chain divides the priority into high, medium, and low according to a preset threshold, wherein the high priority corresponds to the body shape coefficient being greater than a first threshold, such as a large fish, and the motion aggressiveness index being greater than a second threshold, such as fast swimming and frequent turning; the medium priority corresponds to the body shape coefficient being greater than the first threshold or the motion aggressiveness index being greater than the second threshold; and the low priority corresponds to the body shape coefficient being less than the first threshold and the motion aggressiveness index being less than the second threshold.

[0073] For example, the body shape coefficient of a shark is 120 kg, which is greater than the first threshold of 50 kg, and the motion aggressiveness index is 1350, which is greater than the second threshold of 800, so the shark is of high priority; the body shape coefficient of a sardine is 0.2 kg, which is less than the first threshold of 50 kg, and the motion aggressiveness index is 50, which is less than the second threshold of 800, so the sardine is of low priority.

[0074] In some embodiments, the target recognition module further includes an environment monitoring submodule and an adaptive adversarial sample generation network submodule, wherein the environment monitoring submodule monitors the water transparency and the light intensity in real time, and when any index exceeds a preset threshold, the adaptive adversarial sample generation network submodule is automatically activated. The adaptive adversarial sample generation network submodule synthesizes training data simulating complex environments such as high turbidity, strong light, and weak light through the adaptive adversarial sample generation network, dynamically updates the model parameters in the target recognition module, thereby continuously optimizing to ensure that the target recognition module maintains high robustness and recognition accuracy in complex dynamic underwater environments, while reducing target recognition delay and improving real-time performance.

[0075] In an optional embodiment, after obtaining the initial image sequence of the underwater environment in the water area to be detected, the method further includes:

[0076] The real-time environment data corresponding to the initial image sequence collected by the environment monitoring sub-module is acquired; if the real-time environment data is greater than a preset threshold, the adaptive adversarial sample generation network sub-module is controlled to generate an adversarial sample corresponding to the real-time environment data; the initial sample corresponding to the target recognition result is combined with the adversarial sample as training data, and the scale pattern perception backbone network sub-module, the double-flow space-time attention network sub-module, the multi-scale feature passivation network sub-module, and the multi-stage serial classification sub-module are trained based on the training data respectively to obtain a training result; and the model parameters in the scale pattern perception backbone network sub-module, the double-flow space-time attention network sub-module, the multi-scale feature passivation network sub-module, and the multi-stage serial classification sub-module are dynamically updated based on the training result.

[0077] It should be noted that the online adversarial sample generation mechanism and the adaptive adversarial sample generation network sub-module can improve the robustness of the model in a complex dynamic water environment. When the water transparency decreases, such as silt accumulation, algae outbreak, or light intensity changes dramatically, such as overcast days, insufficient light in deep water layers, the initial image sequence will have problems such as blurring, noise, and unbalanced contrast, directly leading to distortion in fish feature extraction, such as scale texture blurring and unclear tail swing trajectory, and further causing misjudgment of the multi-stage serial classification sub-module, such as misclassifying fish as background or confusing similar species.

[0078] The online adversarial sample generation mechanism and the adaptive adversarial sample generation network sub-module automatically generate adversarial samples that fit the current environmental interference, dynamically optimize model parameters, and ensure that the multi-stage serial classification sub-module can still maintain high-precision recognition in a complex environment, while avoiding recognition delays caused by environmental mutations without the need for manual data re-collection and training.

[0079] Specifically, the environment monitoring sub-module collects water transparency and light intensity data in real time, and determines whether the water transparency and light intensity data exceed the preset threshold, such as a water transparency corresponding to a preset threshold of 1 meter or a light intensity corresponding to a preset threshold of 300 lux. If the real-time environment data is greater than the preset threshold, an instruction is sent to the adaptive adversarial sample generation network sub-module to generate an adversarial sample corresponding to the real-time environment data; the generated adversarial sample is combined with the initial sample corresponding to the target recognition result as training data for training, and the model parameters in the scale pattern perception backbone network sub-module, the double-flow space-time attention network sub-module, the multi-scale feature passivation network sub-module, and the multi-stage serial classification sub-module are updated based on the training result until the recognition accuracy of the model parameters on the new sample returns to a stable level.

[0080] For example, when the water body suddenly becomes turbid, the scale texture and tail fin trajectory in the initial image sequence become blurred, which can cause distortion of the spatio-temporal fusion feature map extracted by the scale perception backbone network submodule; and the species fine-grained recognition in the multi-stage serial classification submodule can be misjudged, such as misjudging crucian carp as common carp. At this time, the adaptive adversarial sample generation network submodule generates adversarial samples in a turbid environment, and the scale perception backbone network submodule and the multi-stage serial classification submodule are retrained based on the adversarial samples, so that the scale perception backbone network submodule learns to extract key features from blurred images, such as distinguishing species by body proportions even if the scales are blurred; and the multi-stage serial classification submodule is retrained based on the adversarial samples, so that the multi-stage serial classification submodule adapts to noisy edge features and avoids misjudgment caused by blurred edges.

[0081] S103, determining the carrier pose data of the intelligent fishing system according to the target pixel coordinates corresponding to the target recognition result and the target image sequence, and obtaining the historical position sequence and the target motion feature of the target fish, and the inertial measurement unit data of the intelligent fishing system.

[0082] In some embodiments, the target positioning module includes a fluid topological visual SLAM submodule, a deep tight-coupling feature enhancement network submodule, and a double-state extended Kalman filter submodule, wherein the fluid topological visual SLAM submodule includes a refraction distortion compensation module, a contour-enhanced feature point tracking module, and a dynamic continuous map construction module.

[0083] To achieve real-time three-dimensional spatial coordinate calculation of a moving target fish, achieve centimeter-level positioning accuracy, predict its short-time motion trajectory, and thus accurately capture the target fish later, in an optional embodiment, the step of determining the carrier pose data of the intelligent fishing system according to the target pixel coordinates corresponding to the target recognition result and the target image sequence can include:

[0084] The refraction distortion compensation module is used to perform distortion compensation processing on the target pixel coordinates and the target image sequence to obtain a distortion-free image sequence; the contour-enhanced feature point tracking module is used to perform contour enhancement and feature point tracking processing on the distortion-free image sequence to obtain a set of static feature points of the underwater environment in the distortion-free image sequence; and the dynamic continuous map construction module is used to construct a target map based on the set of static feature points, and determine the carrier pose data according to the target map.

[0085] It should be noted that the intelligent fishing system starts the fluid topological visual SLAM submodule, analyzes feature points such as fish body contours and underwater terrain in the continuous underwater image sequence, i.e., the target image sequence, constructs an underwater environment map in real time, and calculates the pose of the intelligent fishing system itself, i.e., the carrier pose data corresponding to the pose data of the intelligent fishing system.

[0086] Specifically, due to the refraction of light from water with a refractive index of about 1.33 into air with a refractive index of about 1 when underwater imaging, the mapping relationship between the actual three-dimensional coordinates of the object and the two-dimensional pixel coordinates of the image deviates from the ideal perspective model, that is, refraction distortion. For example, the underwater object may appear position offset, shape stretching or proportion distortion in the target image sequence, directly affecting the calculation of the physical position of the feature point. Based on this, the refraction distortion compensation module is a mathematical model constructed based on the refraction law, which performs distortion compensation processing through the real-time refractive index of the water body, the relative depth and angle of the intelligent fishing system and the water surface or the target fish, and corrects the distortion to obtain a non-distortion image sequence, thereby solving the misplacement problem of underwater imaging and ensuring that the feature point coordinates in the image can accurately reflect the actual position in the three-dimensional space.

[0087] Specifically, due to the refraction of light from water with a refractive index of about 1.33 into air with a refractive index of about 1 when underwater imaging, the mapping relationship between the actual three-dimensional coordinates of the object and the two-dimensional pixel coordinates of the image deviates from the ideal perspective model, that is, refraction distortion. For example, the underwater object may appear position offset, shape stretching or proportion distortion in the target image sequence, directly affecting the calculation of the physical position of the feature point. Based on this, the refraction distortion compensation module is a mathematical model constructed based on the refraction law, which performs distortion compensation processing through the real-time refractive index of the water body, the relative depth and angle of the intelligent fishing system and the water surface or the target fish, and corrects the distortion to obtain a non-distortion image sequence, thereby solving the misplacement problem of underwater imaging and ensuring that the feature point coordinates in the image can accurately reflect the actual position in the three-dimensional space.

[0088] Specifically, the dynamic continuous map construction module constructs a target map based on the static feature point set, provides a global reference for the carrier pose, such as constraining the current pose through the known points in the target map, and determines the carrier pose data according to the target map. The dynamic continuous map construction module can eliminate the feature points belonging to dynamic objects in the map based on the output of the target recognition module, such as the category and position of the dynamic target; adopts a sliding window optimization to fuse the newly added feature points with the historical map, corrects the cumulative error, and updates the topology of the map in real time to ensure the continuity of the map when the carrier moves.

[0089] S104, the control target positioning module performs filtering processing based on the carrier pose data, the inertial measurement unit data, the historical position sequence and the target motion feature to determine the target positioning result and the target prediction trajectory result of the target fish.

[0090] In some embodiments, the deep tight coupling feature enhancement network sub-module includes a three-dimensional observation vector construction module, a motion feature extraction module, and a target kinematics constraint module, and the double-state extended Kalman filtering sub-module includes a body state predictor, a target trajectory predictor, and a confidence evaluation calibration module.

[0091] In an optional embodiment, the control three-dimensional observation vector construction module fuses the target pixel coordinates with the inertial measurement unit data to obtain a three-dimensional observation vector of the target fish in the coordinate system of the intelligent fishing system; the motion feature extraction module extracts relative motion features of the target fish and the intelligent fishing system and motion trend features of the intelligent fishing system from the three-dimensional observation vector; the target kinematics constraint module performs kinematics constraint processing on the relative motion features and the motion trend features to obtain depth information of the target fish and the intelligent fishing system; the control body state predictor filters the carrier pose data and the inertial measurement unit data to obtain target carrier pose data, and converts the target pixel coordinates into three-dimensional coordinates through the target carrier pose data to obtain an initial positioning result; the control target trajectory predictor performs filtering processing based on the initial positioning result, a historical position sequence, and target motion features to obtain an initial predicted trajectory result; the confidence evaluation calibration module performs confidence evaluation calibration processing on the initial positioning result and the initial predicted trajectory result to obtain a target positioning result and a target predicted trajectory result.

[0092] Specifically, the target pixel coordinates are specific positions of each target in an image output by the target recognition module after processing a target image sequence. The inertial measurement unit is a sensor integrating an accelerometer and a gyroscope, which can output angular velocity and linear acceleration. Among them, the target pixel coordinates can reflect its horizontal and vertical directions relative to the carrier, such as the target being on the left side of the image, indicating being in the front left of the carrier; the inertial measurement unit data can reflect the motion state of the carrier, such as the target pixel change rate in the image can be deduced when the carrier accelerates forward; the three-dimensional observation vector construction module combines two-dimensional target pixel coordinates with inertial measurement unit data through a mathematical model, which can solve the three-dimensional coordinates of the target in the carrier coordinate system, forming a three-dimensional observation vector of the target fish in the coordinate system of the intelligent fishing system.

[0093] Specifically, the motion feature extraction module can extract relative motion features of the target relative to the carrier, such as relative speed, relative acceleration, motion direction angle, etc.; the motion feature extraction module extracts motion trend features of the carrier itself, such as motion stability and acceleration change rate of the carrier. In addition, the motion feature extraction module can also filter noise and eliminate outliers in the three-dimensional observation vector, such as false pixel coordinates caused by image blur and sudden data caused by vibration.

[0094] Specifically, the target kinematic constraint module utilizes the fish tail swing frequency, turning angular velocity and other characteristic constraints to constrain the motion direction, i.e., the target kinematic constraint module performs kinematic constraint processing on the relative motion characteristics and motion trend characteristics, thereby obtaining the depth information of the target fish and the intelligent fishing system. For example, fish with high tail swing frequency usually have a fast motion speed, and their displacement range in a short time can be constrained; fish with large turning angular velocity have a motion trajectory that deviates to the inside of the turn, and the error range of the depth estimation can be reduced. By limiting the possible motion state of the target through kinematic characteristics and excluding unreasonable depth assumptions through the motion law of the target, the multi-solution of monocular vision is converted into a unique solution, effectively eliminating the depth ambiguity of monocular vision, thereby improving the estimation accuracy of the relative distance of the target and obtaining the depth information of the target fish and the intelligent fishing system.

[0095] Specifically, the body state predictor, as the first level filter core, can suppress the high-frequency jitter of the carrier motion, such as the instantaneous deviation caused by water flow impact, and output stable carrier pose data to provide a reliable observation reference coordinate system for target positioning. Among them, the carrier pose data and the inertial measurement unit data are taken as the input data of the body state predictor, the stable target carrier pose data is obtained, and the target pixel coordinates of the target carrier pose data are converted into three-dimensional coordinates to obtain an initial positioning result with a precision of centimeters.

[0096] Specifically, the target trajectory predictor, as the second level filter core, can establish an adaptive kinematic model based on the three-dimensional coordinates output by the body state predictor, combine the historical trajectory and motion characteristics such as fish tail swing frequency and turning angular velocity, and predict the future trajectory. Among them, the initial positioning result, the historical position sequence and the target motion characteristics are taken as the input data of the target trajectory predictor, and the initial prediction trajectory result is obtained through filtering processing.

[0097] Specifically, the confidence assessment calibration module can dynamically assess the reliability of positioning and trajectory prediction, adjust the noise covariance of the two-level filters to calibrate the error, and avoid the result deviation caused by environmental interference such as turbid water or low recognition confidence. Among them, the confidence assessment calibration module performs confidence assessment calibration processing on the initial positioning result and the initial prediction trajectory result, obtains the target positioning result and the target prediction trajectory result with a precision of centimeters, thereby solving the problems of large visual noise and nonlinear target motion in the underwater environment, and realizing centimeter-level positioning and high-precision trajectory prediction.

[0098] S105, based on the target recognition result, the target positioning result and the target prediction trajectory result, controlling the harpoon launching device to release the harpoon for fishing operation.

[0099] From the above, the embodiment obtains a target image sequence of an underwater environment in a water area containing fish to be detected; performs feature extraction and feature classification processing based on the target image sequence in sequence to determine a target recognition result of a target fish; determines carrier pose data of an intelligent fishing system according to target pixel coordinates corresponding to the target recognition result and the target image sequence, and obtains historical position sequence and target motion features of the target fish, and inertial measurement unit data of the intelligent fishing system; performs filtering processing based on the carrier pose data, the inertial measurement unit data, the historical position sequence and the target motion features to determine a target positioning result and a target prediction trajectory result of the target fish; and controls a harpoon launching device to release a harpoon for fishing operation based on the target recognition result, the target positioning result and the target prediction trajectory result, so as to improve the accuracy of fishing.

[0100] According to the method described in the foregoing embodiments, the following will be further described in detail by way of examples. Please refer to Figure 2 , Figure 2 is a second flowchart of the intelligent fishing method provided by the embodiment of the present application. The specific process of the intelligent fishing method can include:

[0101] S201, obtaining a target image sequence of an underwater environment in a water area containing fish to be detected.

[0102] In the embodiment, the specific description of steps S201-S204 can refer to the description of steps S101-S104 in the foregoing embodiments, which will not be repeated here.

[0103] S202, controlling a target recognition module to perform feature extraction and feature classification processing based on the target image sequence in sequence to determine a target recognition result of a target fish.

[0104] S203, determining carrier pose data of an intelligent fishing system according to target pixel coordinates corresponding to the target recognition result and the target image sequence, and obtaining historical position sequence and target motion features of the target fish, and inertial measurement unit data of the intelligent fishing system.

[0105] S204, controlling a target positioning module to perform filtering processing based on the carrier pose data, the inertial measurement unit data, the historical position sequence and the target motion features to determine a target positioning result and a target prediction trajectory result of the target fish.

[0106] S205, obtaining target dynamic data of the target fish, and visual quality evaluation data corresponding to the target image sequence.

[0107] The target dynamic data includes three-dimensional coordinates, velocity vectors, and predicted trajectories of the target fish, directly reflects the motion state of the target, and is the core basis for decision-making. The visual quality evaluation data is obtained by evaluating the quality of the target image through indicators such as contour sharpness, texture contrast, and motion blur. For example, image blurring caused by high turbidity of the water body will reduce the reliability of the data, thereby providing a reference for subsequent feature weighting.

[0108] In S206, the target decision module performs deep learning processing based on the target dynamic data, the visual quality evaluation data, and the inertial measurement unit data, and determines a first fishing strategy for capturing the target fish.

[0109] In order to obtain the optimal launch timing and angle for the target fish in real time, and ensure that the harpoon hits the key part of the target along the optimal trajectory, in some embodiments, an underwater launch real-time decision method based on environmental adaptation is adopted. The target decision module includes a spatio-temporal adaptive fusion network submodule and a decision-driven deep reinforcement learning submodule. The decision-driven deep reinforcement learning submodule includes a hierarchical state space module and a decision reinforcement strategy generation network module.

[0110] In an optional embodiment, the target dynamic data, the visual quality evaluation data, and the inertial measurement unit data are fused to construct multi-source input data. The spatio-temporal adaptive fusion network submodule extracts features from the multi-source input data in the time dimension and the space dimension to determine first target motion features and first carrier disturbance features, and performs weighted processing on the first target motion features and the first carrier disturbance features to obtain a spatio-temporal fusion feature vector. The hierarchical state space module processes the feature vector in layers to obtain a layered processing result. The decision reinforcement strategy generation network module generates a target decision vector based on the layered processing result, and determines the first fishing strategy based on the target decision vector.

[0111] Specifically, the inertial measurement unit data includes angular velocity, linear acceleration, and other motion parameters of the underwater carrier, reflecting the disturbance state of the carrier itself, such as shaking caused by water flow impact. Through data standardization processing, the target dynamic data, the visual quality evaluation data, and the inertial measurement unit data are associated and mapped to the same feature space to form a structured input matrix, i.e., to construct multi-source input data, thereby laying a foundation for subsequent fusion processing.

[0112] Specifically, the spatio-temporal adaptive fusion network submodule performs feature fusion on the input multi-source input data, dynamically allocates weights through a multi-modal attention mechanism: high weights are assigned to target motion features such as velocity vectors and trajectory change rates to ensure that the decision focuses on the real-time dynamics of the target and determines the first target motion feature; the weights of the carrier disturbance features such as the carrier shaking amplitude reflected by the inertial measurement unit data are adjusted according to the environmental disturbance intensity to obtain the first carrier disturbance feature, such as increasing the weight in strong water flow to preferentially compensate for carrier disturbance; combined with the time dimension, i.e., the time sequence change of the target motion, and the spatial dimension, i.e., the relative position of the target and the carrier, the comprehensive feature vector containing spatio-temporal correlation information is fused to eliminate data redundancy and highlight key features, thereby obtaining a spatio-temporal fused feature vector.

[0113] Specifically, the hierarchical state space module is to perform hierarchical abstraction on the spatio-temporal fused feature vector, decompose the high-dimensional and complex environmental state into high-level task states and low-level detail states, and obtain hierarchical processing results, thereby reducing the decision difficulty of reinforcement learning. The high-level task state contains state information such as whether the target enters the effective range, whether the target motion trend is stable, and whether the carrier is in a low disturbance state; the low-level detail state contains state information such as the target relative distance, real-time velocity vector, carrier instantaneous attitude angle, and visual feature clarity quantization value.

[0114] Specifically, the decision reinforcement strategy generation network module takes the feature vector corresponding to the hierarchical processing result as input, solves it through a multi-layer neural network, and outputs a specific decision vector, i.e., a target decision vector. It should be noted that the target decision vector includes a launch timing parameter and a trajectory parameter, and the first fishing strategy can be determined based on the launch timing parameter and the trajectory parameter.

[0115] Optionally, the optimal launch window is quantified by a launch probability, such as a launch probability greater than 0.8 being determined as a suitable launch timing. The launch probability is combined with factors such as target motion stability, such as a higher probability when moving in a straight line, and carrier attitude stability to generate the launch timing parameter; an optimization algorithm is used to ensure that the intersection of the trajectory and the target predicted trajectory result is the key part of the target, such as the middle of the fish body, while taking into account the deflection of the water flow on the harpoon, such as fine-tuning the angle to compensate for the water flow thrust when following the flow, thereby obtaining the trajectory parameter.

[0116] S207, based on the target recognition result, the target positioning result, and the target predicted trajectory result, controlling the harpoon launching device to release the harpoon for fishing operations according to the first fishing strategy.

[0117] In some embodiments, the target decision module further includes a multi-dimensional reward function submodule, and the decision-driven deep reinforcement learning submodule further includes a decision reinforcement strategy evaluation module. The harpoon launching device can be arranged at the bottom of the carrier, i.e., the intelligent fishing system.

[0118] After the fish spear launching device is controlled to release the fish spear to perform the fishing operation according to the first fishing strategy, a historical fishing result of the fishing operation performed according to the first fishing strategy can be obtained; the historical fishing result is analyzed and processed in multiple dimensions based on the multi-dimensional reward function submodule to obtain a reward result; the decision-making reinforcement strategy evaluation module evaluates the first fishing strategy based on the hierarchical processing result and the reward result to obtain an evaluation result; if the evaluation result meets a preset condition, the first fishing strategy is optimized, and the decision-making reinforcement strategy generation network module generates an updated first fishing strategy.

[0119] It should be noted that the target decision module ensures the adaptability of the decision through strategy evaluation and online optimization. Specifically, in order to evaluate the effectiveness of the strategy, the multi-dimensional reward function submodule evaluates the decision effect from multiple dimensions, including hit accuracy, i.e., whether the target key part is hit, low disturbance, i.e., the degree of influence on non-target organisms, launch efficiency, i.e., whether the launch is in the optimal window, etc., and the reward value is calculated to obtain a reward result. The higher the reward value, the better the strategy.

[0120] Specifically, the hierarchical state space module divides the environment state such as turbidity and water flow speed, the target state such as the motion mode, and the carrier state such as the attitude stability into different levels, and evaluates the adaptability of the strategy in a specific state layer by layer, such as whether the strategy in a high turbidity state can still maintain accuracy, to obtain a hierarchical processing result. The decision-making reinforcement strategy evaluation module evaluates the first fishing strategy based on the hierarchical processing result and the reward result to obtain an evaluation result.

[0121] Specifically, if the reward value is lower than a preset threshold, i.e., the strategy effect is poor or the environment state changes significantly, such as from low turbidity to high turbidity, the online meta-learning optimizer is triggered, the parameters of the decision-making reinforcement strategy generation network module are updated based on the latest operation data such as the missed case and the environment parameter change record, to realize the optimization of the first fishing strategy, so that the updated first fishing strategy quickly adapts to the new environment, and ensures that the fish spear can always hit the target with high accuracy and low delay.

[0122] In some optional embodiments, please refer to Figure 3 , Figure 3 is a third flowchart of an intelligent fishing method provided by the embodiments of the present application. The specific process of the intelligent fishing method can include:

[0123] S301, obtaining a target image sequence of an underwater environment in a water area containing fish to be detected.

[0124] In this embodiment, the specific description of steps S301-S304 can be referred to the description of steps S101-S104 in the above embodiments, which will not be repeated here.

[0125] S302, the control target recognition module sequentially performs feature extraction and feature classification processing based on the target image sequence to determine a target recognition result of the target fish.

[0126] S303, determining the carrier pose data of the intelligent fishing system according to the target pixel coordinates corresponding to the target recognition result and the target image sequence, and obtaining the historical position sequence and the target motion feature of the target fish, and the inertial measurement unit data of the intelligent fishing system.

[0127] S304, the control target positioning module performs filtering processing based on the carrier pose data, the inertial measurement unit data, the historical position sequence and the target motion feature to determine the target positioning result and the target prediction trajectory result of the target fish.

[0128] S305, obtaining the target state data of the target fish, the current carrier pose data of the intelligent fishing system and the environmental disturbance data, wherein the target state data includes the target three-dimensional coordinates corresponding to the target positioning result, the target prediction trajectory corresponding to the target prediction trajectory result and the relative speed vector of the target fish and the intelligent fishing system.

[0129] The target state data includes the three-dimensional coordinates of the target fish, i.e. the target three-dimensional coordinates corresponding to the target positioning result, the relative speed vector of the target fish and the intelligent fishing system, i.e. the speed difference between the target fish and the intelligent fishing system, and the target prediction trajectory corresponding to the target prediction trajectory result, which directly reflects the motion trend of the target fish.

[0130] The current carrier pose data is the real-time pose of the water-borne carrier, i.e. the carrier's own three-dimensional coordinates and attitude angle, which serves as the reference for attitude adjustment.

[0131] The environmental disturbance data is the flow disturbance coefficient, which can quantify the impact force of the flow on the carrier and is used to compensate for environmental disturbances.

[0132] S306, the control target decision module adjusts the carrier pose of the intelligent fishing system based on the target state data, the current carrier pose data and the environmental disturbance data to determine a second fishing strategy for capturing the target fish.

[0133] To achieve accurate aiming of the harpoon launching device by adjusting the attitude of the water-borne carrier, in an optional embodiment, a body adaptive attitude control method based on a double-branch spatio-temporal encoder is adopted. The target decision module further includes a double-branch spatio-temporal encoding submodule, an attitude calculation output network submodule and a turbine power distribution network submodule, wherein the double-branch spatio-temporal encoding submodule includes a target motion feature branch module, a carrier disturbance feature branch module and an environmental attention fusion module.

[0134] In some embodiments, the target state data is feature extracted based on the target motion feature branch module to obtain second target motion features; the current carrier pose data and the environmental disturbance data are feature extracted based on the carrier disturbance feature branch module to obtain second carrier disturbance features; the second target motion features and the second carrier disturbance features are weighted and fused by the environmental attention fusion module to obtain a first fusion feature vector; the first fusion feature vector is processed by the attitude solution output network submodule to obtain first angle control data corresponding to the intelligent fishing system; the first power distribution matrix is generated based on the first angle control data by the control turbine power distribution network submodule; and the turbine in the intelligent fishing system is powered by the first power distribution matrix to obtain a second fishing strategy.

[0135] Specifically, the target motion feature branch module extracts the motion features of the target, such as the rate of change of speed, the trajectory curvature, and the relative distance change trend, by combining convolutional neural networks with recurrent neural networks, to form a target dynamic feature vector, i.e., the second target motion features.

[0136] Specifically, the carrier disturbance feature branch module analyzes the fluctuations in the real-time pose of the unmanned aerial vehicle, such as the attitude angle jitter and the water flow disturbance coefficient caused by the water flow, learns the motion disturbance patterns of the carrier, and outputs a carrier disturbance feature vector, i.e., the second carrier disturbance features.

[0137] Specifically, the environmental attention fusion module dynamically adjusts the weights of environmental parameters such as turbidity and illumination by weighted fusion processing of the second target motion features and the second carrier disturbance features, forms a comprehensive feature matrix containing spatio-temporal correlations, i.e., the first fusion feature vector, and provides a basis for attitude solution. For example, when the turbidity is high, the target motion features need to be given a higher weight because the visual data is less reliable; when the illumination changes suddenly, the carrier stability features need to be given a higher weight because the carrier's own attitude needs to be kept stable.

[0138] Specifically, the attitude solution output network submodule processes the first fusion feature vector by multilayer perceptron and attitude dynamics model operation to obtain first angle control data corresponding to the intelligent fishing system. The first angle control data includes pitch angle, yaw angle, and roll compensation angle. The pitch angle is used to adjust the inclination angle of the carrier vertically to control the vertical alignment of the harpoon launcher; the yaw angle is used to control the horizontal turning of the carrier to ensure the horizontal alignment of the harpoon launcher to the target; and the roll compensation angle is used to offset the lateral inclination of the carrier caused by the water flow to maintain the horizontal stability of the harpoon launcher. The attitude solution output network submodule is trained by reinforcement learning to make the output angle control data match the target predicted trajectory in real time, ensuring the consistency of the harpoon axis and the target motion direction.

[0139] Specifically, the control mode of the turbine power distribution network submodule on the turbine power distribution is that the turbine power distribution network submodule calculates and outputs a first power distribution matrix, which determines the size and direction of the thrust of the turbine, such as the left turbine force to realize right turning, and the lower turbine force to adjust the pitch angle. The control mode of the turbine power distribution network submodule on the turbine cooperative control is that the turbine power distribution network submodule divides the functions of the turbine, forms the torque through differentiated power output, and quickly responds to the attitude control demand. For example, when the yaw angle needs to be adjusted, the left and right turbines produce a thrust difference, drive the carrier body to rotate horizontally, and realize precise steering. The turbine power distribution network submodule controls the carrier pose through turbine power distribution and turbine cooperative control, obtains a second fishing strategy, and thus realizes precise aiming of the harpoon launching device.

[0140] In some embodiments, the target decision module further includes a feedback compensation network submodule. After the turbine in the intelligent fishing system is powered by the power distribution matrix to obtain the second fishing strategy, the first carrier pose data of the carrier after the carrier pose adjustment by the intelligent fishing system through the second fishing strategy can be obtained. Based on the analysis and processing of the first carrier pose data and the power distribution matrix by the feedback compensation network submodule, the angle deviation compensation data is obtained. The turbine power distribution network submodule adjusts the first angle control data based on the angle deviation compensation data to generate a second power distribution matrix. The turbine in the intelligent fishing system is powered by the second power distribution matrix to obtain an updated second fishing strategy.

[0141] Specifically, the feedback compensation network submodule obtains real-time carrier pose data after the carrier pose adjustment by the intelligent fishing system through the second fishing strategy, that is, the first carrier pose data, obtains the angle deviation compensation data by calculating the deviation between the current attitude angle and the target attitude angle, such as the pitch angle error, and then adjusts the power distribution matrix in the opposite direction to correct the turbine thrust, so as to offset the influence of factors such as water flow disturbance and turbine response delay, ensure the rapid and accurate alignment of the harpoon axis to the target, and improve the launching accuracy.

[0142] S307, based on the target recognition result, the target positioning result, and the target prediction trajectory result, controlling the harpoon launching device to release the harpoon for fishing operation according to the first fishing strategy and / or the second fishing strategy.

[0143] In some optional embodiments, please refer to Figure 4 , Figure 4 is a fourth flowchart of the intelligent fishing method provided by the embodiments of the present application. The specific process of the intelligent fishing method can include:

[0144] S401, obtaining a target image sequence of an underwater environment in a water area containing fish to be detected.

[0145] In this embodiment, the specific description of steps S401-S404 can refer to the description of steps S101-S104 in the above embodiment, which will not be repeated here.

[0146] S402, the target recognition module sequentially performs feature extraction and feature classification processing based on the target image sequence to determine the target recognition result of the target fish.

[0147] S403, the target recognition result corresponding to the target pixel coordinates and the target image sequence determines the carrier pose data of the intelligent fishing system, and the historical position sequence and the target motion feature of the target fish, and the inertial measurement unit data of the intelligent fishing system.

[0148] S404, the target positioning module performs filtering processing based on the carrier pose data, the inertial measurement unit data, the historical position sequence and the target motion feature to determine the target positioning result and the target prediction trajectory result of the target fish.

[0149] S405, the target execution module sequentially performs data fusion and precision optimization processing based on the target three-dimensional coordinates, the target prediction trajectory and the inertial measurement unit data to generate the rudder driving signal.

[0150] To achieve accurate fishing of the target fish, in an optional embodiment, a high-precision intelligent rudder driving method is adopted. The target execution module includes a self-calibration feature extraction network sub-module, a double-branch precision control network sub-module, and a multi-stage control instruction generation sub-module. The double-branch precision control network sub-module includes a timing precision control branch module, a disturbance suppression branch module, and a double-path collaborative fusion module. The multi-stage control instruction generation sub-module includes a microsecond-level trigger module, a super-precision angle control module, and a dynamic phase compensation module.

[0151] In some embodiments, the self-calibration feature extraction network sub-module performs weighted fusion processing on the target three-dimensional coordinates, the target prediction trajectory, and the inertial measurement unit data to obtain a second fusion feature vector. The historical rudder driving signal of the rudder in the intelligent fishing system and the current underwater environment data in the water area to be detected are obtained. The timing precision control branch module performs deviation analysis processing on the historical rudder driving signal to obtain control deviation compensation data. The disturbance suppression branch module performs weighted processing on the current environment data to obtain environmental disturbance compensation data. The double-path collaborative fusion module performs compensation and fusion processing on the second fusion feature vector based on the control deviation compensation data and the environmental disturbance compensation data to obtain a target fusion feature vector. The microsecond-level trigger module, the super-precision angle control module, and the dynamic phase compensation module perform precision optimization processing on the target fusion feature vector respectively to obtain the rudder driving signal.

[0152] Specifically, the self-calibration feature extraction network submodule fuses multi-source data, receives target three-dimensional coordinates, target predicted trajectory, and inertial measurement unit data, eliminates data bias through a self-calibration algorithm, and generates a standardized feature vector containing target position, motion trend, and carrier state, i.e., a second fusion feature vector, providing accurate input for subsequent control decisions.

[0153] The dual-branch precision control network submodule adopts a dual-branch parallel processing architecture to optimize control precision. Specifically, the timing precision control branch module learns historical control bias, such as the rudder angle error of the past 10 launches, establishes an error compensation model, predicts the bias that may be generated by the current control instruction, and makes corrections in advance. For example, if historical data shows that the rudder has a positive 0.1 degree offset at a certain angle, the real-time instruction automatically subtracts the offset. The disturbance suppression branch module dynamically weights environmental interference factors such as water flow speed and water resistance based on the attention mechanism, giving high weight to strong interference scenarios and prioritizing the output of anti-disturbance control parameters, i.e., environmental interference compensation data, such as increasing the rudder driving torque to offset the water flow impact. The dual-path collaborative fusion module integrates control bias compensation data and environmental interference compensation data, balances timing precision and anti-disturbance requirements, and outputs an optimized control feature vector, i.e., a target fusion feature vector.

[0154] The multi-stage control instruction generation submodule generates network output precision control signals. Through the synergistic action of three-level modules, the rudder is controlled with high precision. Specifically, the microsecond-level trigger module generates rudder driving signals at microsecond-level response speed after receiving the launch instruction, controls the rudder to start, ensures that the buckle unlocking time completely matches the optimal launch window predicted by the target trajectory, and controls the time jitter within microsecond level. The ultra-precision angle control module superimposes a dynamic correction term calculated based on water flow disturbance and carrier attitude deviation on the basic control angle, achieving sub-angle level resolution, such as angle control precision less than 0.1 degree, ensuring the accuracy of the buckle rotation angle. The dynamic phase compensation module analyzes the synchronization error between the rudder mechanical response delay and the control instruction, generates phase compensation parameters in real time, compresses the synchronization error between the instruction and the execution to within milliseconds, and avoids launch time deviation caused by delay.

[0155] In some embodiments, the target execution module further comprises a closed-loop precision feedback network submodule, wherein the closed-loop precision feedback network submodule comprises an error compensation model and a high-precision detection encoder. After the microsecond-level trigger module, the ultra-precision angle control module, and the dynamic phase compensation module are used to respectively perform precision optimization processing on the target fusion feature vector, the actual control data of the rudder collected by the high-precision detection encoder can be obtained; the actual control data and the target control data corresponding to the rudder driving signal are compared to determine the control error result; the second fusion feature vector is weighted and compensated by the error compensation model according to the control error result to obtain a third fusion feature vector after compensation; and an updated rudder driving signal is generated based on the third fusion feature vector.

[0156] The closed-loop precision feedback network submodule can optimize the control parameters in real time. Specifically, the high-precision detection encoder collects the actual rotation angle data of the rudder in real time, compares the target control angle, calculates the angle error, and thus determines the control error result. The error compensation model is used to analyze the error sources, such as the resistance changes caused by spring fatigue and the influence of water temperature on the performance of the rudder, to dynamically adjust the weight parameters of the double-branch control network, such as increasing the historical deviation compensation weight to correct long-term drift, to obtain a third fusion feature vector after compensation, and to generate an updated rudder driving signal based on the third fusion feature vector. It should be noted that the closed-loop precision feedback network submodule can achieve microsecond-level time precision and sub-angle-level spatial precision for rudder driving, ensuring accurate alignment of the fish spear launch direction and the target trajectory.

[0157] S406, based on the rudder driving signal, controlling the fish spear launching device to release the fish spear for fishing operation.

[0158] In some optional embodiments, please refer to Figure 5 , Figure 5 is the fifth flowchart of the intelligent fishing method provided by the embodiments of the present application. The specific process of the intelligent fishing method can include:

[0159] S501, obtaining a target image sequence of an underwater environment in a water area containing fish to be detected.

[0160] In this embodiment, the specific description of steps S501-S504 can be referred to the description of steps S101-S104 in the above embodiments, which will not be repeated here.

[0161] S502, controlling the target identification module to sequentially perform feature extraction and feature classification processing based on the target image sequence to determine a target identification result of the target fish.

[0162] S503, determine the carrier pose data of the intelligent fishing system according to the target pixel coordinates corresponding to the target recognition result and the target image sequence, and obtain the historical position sequence and target motion characteristics of the target fish, and the inertial measurement unit data of the intelligent fishing system.

[0163] S504, control the target positioning module to perform filtering processing based on the carrier pose data, the inertial measurement unit data, the historical position sequence, and the target motion characteristics, to determine the target positioning result and the target prediction trajectory result of the target fish.

[0164] S505, control the harpoon launching device to release the harpoon for fishing operation.

[0165] S506, obtain the instantaneous image corresponding to the target fishing operation.

[0166] To continuously improve the success rate of target capture in complex water areas, in an optional embodiment, a fishing operation effect self-evolution learning method is adopted.

[0167] The target fishing operation is any fishing operation that has occurred. After the target fishing operation is completed, the visual sensor of the water-borne carrier immediately captures the picture of the fishing moment including the harpoon hit state, the target fish reaction, the surrounding environment, etc. as the instantaneous image, which is used as the original data for operation effect evaluation.

[0168] S507, analyze and process the instantaneous image to obtain a fishing efficiency index, and optimize the target positioning result and the harpoon launching strategy of the harpoon launching device based on the fishing efficiency index to obtain an optimized control strategy.

[0169] The target evolution module includes a multi-branch deduction evaluation network submodule, a fishing efficiency index fusion network submodule, and a double closed-loop collaborative optimization network submodule. The double closed-loop collaborative optimization network submodule includes an identification and positioning closed-loop module and a decision-making and propulsion collaborative closed-loop module.

[0170] In some embodiments, the multi-branch deduction evaluation network submodule is used to analyze and process the instantaneous image to obtain an evaluation result of the target fishing operation; the fishing efficiency index fusion network submodule is used to perform weighted fusion processing on the evaluation result to obtain a fishing efficiency index; and the double closed-loop collaborative optimization network submodule is used to perform positioning optimization on the target positioning result and strategy optimization on the harpoon launching strategy based on the fishing efficiency index to obtain an optimized control strategy.

[0171] The multi-branch deduction evaluation network submodule analyzes the instantaneous image based on the hit analysis module, the vitality evaluation module, and the behavior prediction module to obtain an evaluation result. Specifically, the hit analysis module corresponds to analyze and calculate the hit probability and the miss distance vector. The hit probability is calculated by image recognition to determine whether the harpoon hits the target, and the hit probability of a single operation is calculated. For example, if the overlap between the hit area and the target contour is greater than 70%, it is determined to be a hit. The miss distance vector is calculated by coordinate comparison to generate a miss distance vector if it is not hit, which quantifies the degree of deviation. The miss distance vector includes the miss distance and direction, such as a deviation of 5 cm to the right of the target.

[0172] Specifically, the vitality evaluation module uses an image segmentation algorithm to extract the key areas of the target fish, such as the gill and spine, which are prone to damage. The hit area ratio is calculated, which is the area ratio of the key area hit by the harpoon, to evaluate the damage degree of the target fish. For example, if the hit area ratio is less than 30%, it is considered to be low damage.

[0173] Specifically, the behavior prediction module analyzes the motion trajectory of the target fish in the instantaneous image, extracts the muscle contraction frequency based on the biomechanical model, and calculates the muscle contraction frequency based on the fish body swing amplitude and speed. The contraction frequency reflects the stress response of the target after being hit, and quantifies the spasm intensity and escape ability. For example, a high contraction frequency indicates a high escape potential, indicating that the next launch opportunity needs to be faster.

[0174] Specifically, after analyzing the instantaneous image by the hit analysis module, the vitality evaluation module, and the behavior prediction module to obtain the evaluation result, the fishing efficiency index fusion network submodule performs weighted fusion processing on the evaluation results of the hit analysis module, the vitality evaluation module, and the behavior prediction module to generate a fishing efficiency index. For example, the fishing efficiency index is valued between 0 and 1, and the higher the fishing efficiency index, the better the operation effect. The weight distribution of the evaluation results is dynamically adjusted according to the scene requirements. For example, in the breeding scene, the low damage weight is higher, and in the ecological protection scene, the hit accuracy weight is higher.

[0175] Specifically, the double closed loop collaborative optimization network submodule optimizes the target positioning result based on the fishing efficiency index. For example, based on the miss distance vector, the anchor box parameters of the target recognition module are corrected online to reduce the subsequent positioning error. For example, if the miss distance vector shows that the target is deviated to the left for several times, the anchor box is shifted to the left for calibration. The double closed loop collaborative optimization network submodule optimizes the harpoon launch strategy based on the fishing efficiency index. For example, based on the damage rate of the vitality evaluation module, the harpoon launch parameters of the target decision module are dynamically adjusted to ensure the fishing efficiency while reducing the target loss. For example, if the damage rate is too high, the launch intensity is reduced or the hit area is adjusted to a non-critical part. Based on the positioning optimization of the target positioning result and the strategy optimization of the harpoon launch strategy, an optimized control strategy is obtained.

[0176] S508, an initial intelligent fishing edge model corresponding to the target fishing operation is obtained, and a large shore-based water-bottom world model is obtained.

[0177] S509, based on the optimization control strategy and the large shore-based water-bottom world model, the target evolution module updates and upgrades the initial intelligent fishing edge model to obtain a target intelligent fishing edge model.

[0178] The target evolution module synchronizes the initial intelligent fishing edge model of the water download body to the large shore-based water-bottom world model in real time, and the large shore-based water-bottom world model is a cloud model with massive data processing and deep learning capabilities. Among them, the large shore-based water-bottom world model can optimize the initial intelligent fishing edge model from the multi-water area migration learning engine, the feature consistency optimization network and the global knowledge distillation compression.

[0179] Specifically, the multi-water area migration learning engine uses historical operation data of different water areas such as freshwater lakes, seawater areas, and high-turbidity ponds to train the intelligent fishing edge model to adapt to diversified environments, such as learning the light scattering law of different water bodies; the feature consistency optimization network aligns the target features in different scenes, such as the feature differences of fish scales in different turbidity water bodies, to ensure the recognition stability of the intelligent fishing edge model in cross-environment; the global knowledge distillation compression distills the complex knowledge learned by the large model, such as high-precision trajectory prediction algorithm, into lightweight model parameters, compresses the data volume, and facilitates edge device storage and update.

[0180] It should be noted that the lightweight edge upgrade package generated by the edge model through training and optimization control strategy optimization in the large shore-based water-bottom world model is pushed to the water download body, the online update of the edge model is realized, the target intelligent fishing edge model is obtained, and the target recognition accuracy, decision efficiency and fishing success rate of the edge model in complex water areas are improved.

[0181] In some embodiments, please refer to Figure 6 , Figure 6 is the sixth flowchart of the intelligent fishing method provided by the embodiments of the present application. The specific process of the intelligent fishing method can include:

[0182] S601, data synchronization and filtering processing are performed on the inertial measurement unit data and the target image sequence through the data synchronization and preprocessing network sub-module to obtain multi-degree-of-freedom data.

[0183] In order to protect the safety of the water download body system and prevent false triggering and realize intelligent interlocking control, in an optional embodiment, an underwater multi-degree-of-freedom intelligent interlocking safety control method is adopted. The intelligent fishing system further includes a safety control module, and the safety control module includes a data synchronization and preprocessing network sub-module.

[0184] Specifically, the Kalman filtering algorithm is used to synchronize the inertial measurement unit data and the target image sequence in real time. The inertial measurement unit data includes three-axis acceleration reflecting the linear motion intensity of the UAV and three-axis angular velocity reflecting the rotational motion intensity, with a total of six degrees of freedom data. The target image sequence corresponds to the visual sensor data, which includes target center coordinates and target size, with a total of three degrees of freedom. The fusion of the inertial measurement unit data and the target image sequence forms a nine-degree-of-freedom data stream, which eliminates the measurement noise of a single sensor and improves the data reliability.

[0185] S602, if the multi-degree-of-freedom data is greater than the preset degree-of-freedom threshold, the intelligent fishing system is marked as faulty by the safety control module, and the intelligent fishing system is locked.

[0186] If any value of the acceleration and angular velocity in the real-time monitored data stream exceeds the preset safety threshold, such as acceleration greater than 5 and angular velocity greater than 10, the intelligent fishing system is marked as faulty by the safety control module, and the intelligent fishing system is locked, triggering subsequent in-depth analysis.

[0187] S603, the multi-degree-of-freedom data is analyzed for feature risk from the time domain, frequency domain, and spatial domain by the time-frequency-space dynamic analysis network submodule, obtaining multi-dimensional risk features.

[0188] In some embodiments, the safety control module further includes a time-frequency-space dynamic analysis network submodule, an attention fusion gate network submodule, a time series memory enhancement network submodule, and a dynamic stability decision network submodule.

[0189] The time-frequency-space dynamic analysis network submodule analyzes the multi-degree-of-freedom data from the time domain, frequency domain, and spatial domain through three branches, capturing the system state comprehensively. Specifically, the time domain analysis branch focuses on the instantaneous changes in the time dimension to extract acceleration pulse features such as peak value and duration of sudden acceleration, and identifies whether there is a violent impact such as the UAV colliding with the water bottom rock.

[0190] Specifically, the frequency domain analysis branch converts the angular velocity data to the frequency domain through Fourier transform, analyzes the angular vibration energy spectrum, identifies abnormal vibration frequencies such as specific frequency vibrations caused by gear failure, and judges whether the mechanical structure is abnormal.

[0191] Specifically, the spatial analysis branch is based on the target center coordinates and the carrier pose to construct a target-carrier coupling matrix, quantize the relative distance and orientation relationship between the target and the carrier, and evaluate the spatial safety risk.

[0192] S604, the multi-dimensional risk features are weighted and fused by the attention fusion gate network submodule to obtain an instantaneous risk feature vector.

[0193] The multi-dimensional risk features obtained by performing feature risk analysis on the multi-degree-of-freedom data from the time domain dimension, the frequency domain dimension, and the spatial dimension are input to an attention fusion gate network submodule, the attention fusion gate network submodule performs weighted fusion processing on the multi-dimensional risk features, dynamically allocates feature weights, and obtains an instantaneous risk feature vector. For example, the attention fusion gate network submodule gives high weights to high-risk features such as a coupling matrix of a close-range target and a high-intensity acceleration pulse, and gives priority to potential dangers; and gives low weights to low-interference features such as a stable angular vibration spectrum, and reduces the influence of redundant information.

[0194] In S605, the time sequence memory enhancement network submodule analyzes and processes the instantaneous risk feature vector to obtain a dynamic risk feature vector.

[0195] Specifically, the time sequence memory enhancement network submodule combines a long short-term memory mechanism to learn the association between historical states and current features, thereby analyzing and processing the instantaneous risk feature vector to generate a state vector containing time sequence association, i.e., a dynamic risk feature vector, which comprehensively reflects the dynamic change process of the system from normal to abnormal, and provides time sequence basis for decision-making. For example, if three consecutive frames of data all show high acceleration pulses, the time sequence memory enhancement network submodule will judge it as a continuous fault rather than a transient disturbance.

[0196] In S606, the dynamic stability decision network submodule performs stability analysis on the dynamic risk feature vector to obtain a stability probability result and a species risk distribution result, and determines to unlock or lock the intelligent fishing system based on the stability probability result and the species risk distribution result.

[0197] Specifically, the dynamic risk feature vector is input to the dynamic stability decision network submodule, and the dynamic stability decision network submodule performs stability analysis to obtain a stability probability result and a species risk distribution result. The stability probability result can take a value of 0-1, and the higher the value, the more stable the system; the species risk distribution result can quantify whether the target is a protected species or is in a forbidden fishing area, etc., to avoid misfishing.

[0198] Optionally, when the stability probability result is greater than or equal to 0.9 and the species risk distribution result meets a safety threshold, it is determined that the unlocking condition is met, and a safety unlocking signal is output to unlock the intelligent fishing system. Correspondingly, when the stability probability result is less than 0.9 and / or the species risk distribution result does not meet the safety threshold, it is determined that the unlocking condition is not met, and a locking signal is output to lock the intelligent fishing system.

[0199] In some embodiments, the safety control module further includes a multi-stage runaway protection network submodule, which triggers the rudder power circuit to lock when the risk value is greater than a threshold, thereby realizing intelligent interlocking control and ensuring the safety of the water-borne vehicle system operation.

[0200] Specifically, the multi-stage failure protection network sub-module calculates a system comprehensive risk value in real time. When the risk value exceeds an emergency threshold of 0.7, the multi-stage failure protection network is triggered. In some embodiments, different protection actions can be taken according to the risk level. For example, when the risk value is 0.5-0.7, it is determined as low risk, the rudder motor power is temporarily locked, the launch is prohibited, and the system is self-checked; for example, when the risk value is greater than 0.7, it is determined as high risk, the rudder motor is immediately disconnected from the main power supply, and the standby buoyancy device is started to make the carrier float to a safe water area to prevent equipment damage or misoperation to expand the risk.

[0201] In some embodiments, referring to Figure 7 , Figure 7 is a seventh flowchart of an intelligent fishing method provided by the embodiments of the present application. The specific process of the intelligent fishing method can include:

[0202] S701, obtaining target image and target inertial measurement unit data corresponding to the time when the fish spear launching device releases the fish spear, and performing fusion processing on the target image and the target inertial measurement unit data through a cross-modal data fusion network sub-module to obtain fusion data.

[0203] In order to solve the problem of disturbance to the water carrier attitude caused by the reaction force generated at the moment of fish spear launching, in some embodiments, a posture self-stabilization control method based on a dynamic feature fusion engine is proposed. The intelligent fishing system further includes a posture control module, and the posture control module includes a cross-modal data fusion network sub-module, a dynamic feature fusion engine sub-module, a motion trajectory integration network sub-module, and an execution optimization network sub-module.

[0204] Specifically, the cross-modal data fusion network sub-module eliminates data heterogeneity and filters high-frequency noise through standardized processing and fusion processing of unified coordinate system and timestamp alignment, generates structured original perception data set, i.e. fusion data, and provides a basis for subsequent feature extraction.

[0205] S702, based on the dynamic feature fusion engine sub-module, the fusion data is subjected to feature analysis and weighted processing respectively to obtain a comprehensive disturbance feature vector.

[0206] Specifically, the dynamic feature fusion engine submodule processes the fusion data for vibration feature extraction, position correlation modeling, environmental disturbance calibration, and adaptive weight allocation. Specifically, vibration feature extraction is based on Fourier transform analysis of the acceleration data of the target inertial measurement unit data to extract the vibration frequency caused by the reaction force, the amplitude reflecting the impact force, and the decay rate to judge the vibration duration, accurately depicting the dynamic characteristics of the reaction force. Position correlation modeling is to construct a spatial coupling model of the carrier and the harpoon launcher combined with the relative position data of the visual sensor, quantifying the moment influence of the launch reaction force on the carrier's center of gravity, providing a spatial coordinate reference for thrust compensation. Environmental disturbance calibration introduces environmental parameters such as water flow velocity and water density, and corrects the disturbance characteristics through a fluid dynamics model to ensure that the feature analysis adapts to the real-time environment. For example, when launching downstream, the water flow will amplify the influence of the reaction force, which needs to be compensated.

[0207] The adaptive weight allocation module dynamically adjusts the weights of the above three types of features according to the disturbance intensity, and fuses to form a comprehensive disturbance feature vector, highlighting the factors that have the greatest impact on attitude stability. For example, when there is strong vibration, the vibration feature weight is increased, and when there is strong water flow, the environmental disturbance weight is increased.

[0208] S703, the historical motion trajectory of the intelligent fishing system before the harpoon launcher releases the harpoon is obtained, and the historical motion trajectory and the comprehensive disturbance feature vector are integrated and analyzed by the motion trajectory integration network submodule to obtain the attitude deviation amount.

[0209] Specifically, the motion trajectory integration network submodule fuses the comprehensive disturbance feature vector with the historical motion trajectory of the carrier, predicts the attitude change trend through time series modeling, and obtains the attitude deviation amount, providing a forward-looking basis for compensation control.

[0210] S704, the real-time compensation thrust vector of the turbine in the intelligent fishing system is calculated based on the attitude deviation amount by executing the optimization network submodule, and the target turbine control parameters are obtained based on the real-time compensation thrust vector, so as to control the turbine through the target turbine control parameters.

[0211] Specifically, the execution optimization network submodule is based on the predicted attitude deviation amount, combined with the dynamics model of the carrier, such as mass distribution and turbine thrust characteristics, to calculate the real-time compensation thrust vector required to offset the reaction torque, and to determine the thrust size and direction of each turbine, to obtain the real-time compensation thrust vector. For example, the left turbine needs to increase the thrust by 20% to offset the roll, and the forward-inclined turbine needs to reverse the force to suppress the pitch.

[0212] The output turbine control parameter: converting the real-time compensation thrust vector into a specific turbine control instruction such as voltage and rotating speed, i.e. the target turbine control parameter, to ensure that the accuracy of the target turbine control parameter reaches the sub-Newton level, i.e. the thrust adjustment accuracy is less than 0.1 N, and meets the millisecond-level response requirement, i.e. the time from the disturbance to the output of the target turbine control parameter is less than 5 ms. Through the coordinated fast response and accurate thrust vector control of the turbine, the influence of the launch reaction force on the roll, pitch and yaw angles of the water-borne body is effectively neutralized, the rapid recovery and high stability of the water-borne body attitude after launching are ensured, and a stable platform foundation is provided for subsequent actions such as observing the launching effect, withdrawing or re-launching.

[0213] In addition, the embodiment of the present application further provides an intelligent fishing device. Please refer to Figure 8 , Figure 8 is a structural schematic diagram of an intelligent fishing device provided by the embodiment of the present application. The intelligent fishing device 800 can be applied to an intelligent fishing system, and the intelligent fishing system at least includes a target identification module, a target positioning module and a harpoon launching device. Wherein, the intelligent fishing device 800 can include an acquisition module 801, a first processing module 802, a second processing module 803, a third processing module 804 and a control module 805, and the details are as follows:

[0214] The acquisition module 801 is used for acquiring a target image sequence of an underwater environment in a water area containing fish to be detected;

[0215] The first processing module 802 is used for controlling the target identification module to sequentially perform feature extraction and feature classification processing based on the target image sequence, and determine a target identification result of the target fish;

[0216] The second processing module 803 is used for determining carrier pose data of the intelligent fishing system according to target pixel coordinates corresponding to the target identification result and the target image sequence, and acquiring historical position sequence and target motion characteristics of the target fish, and inertial measurement unit data of the intelligent fishing system;

[0217] The third processing module 804 is used for controlling the target positioning module to perform filtering processing based on the carrier pose data, the inertial measurement unit data, the historical position sequence and the target motion characteristics, and determine a target positioning result and a target prediction trajectory result of the target fish;

[0218] The control module 805 is used for controlling the harpoon launching device to release a harpoon for fishing operation based on the target identification result, the target positioning result and the target prediction trajectory result.

[0219] Optionally, the target recognition module comprises an underwater optical compensation algorithm submodule, and the acquisition module 801 is further configured to: acquire an initial image sequence of an underwater environment in the water area to be detected; and perform despeckling processing on the initial image sequence based on the optical compensation algorithm submodule to eliminate turbid medium scattering noise in the initial image sequence, to obtain a target image sequence.

[0220] Optionally, the target recognition module further comprises a scale perception backbone network submodule, a dual-stream spatiotemporal attention network submodule, a multi-scale feature passivation network submodule, and a multi-stage serial classification submodule, wherein the dual-stream spatiotemporal attention network submodule comprises a spatial feature stream attention module, a temporal feature stream attention module, and a dual-stream feature integration module, and the first processing module 802 is further configured to: perform feature extraction on the target image sequence from a spatial dimension and a temporal dimension respectively based on the scale perception backbone network submodule, to obtain a spatiotemporal fusion feature map of the fish; control the spatial feature stream attention module to perform weighted processing on the spatiotemporal fusion feature map in the spatial dimension, to obtain a target morphological feature of the fish; control the temporal feature stream attention module to obtain a plurality of motion features of the spatiotemporal fusion feature map in the temporal dimension, to obtain a target motion trajectory feature of the fish; perform cross-dimension feature integration on the target morphological feature and the target motion trajectory feature based on the dual-stream feature integration module, to obtain an initial feature vector of the fish; control the multi-scale feature passivation network submodule to sequentially perform multi-scale feature extraction and feature passivation processing on the initial feature vector, to compress background feature interference, to obtain a target feature vector; and perform multi-stage feature classification on the target feature vector based on the multi-stage serial classification submodule, to determine a target recognition result of the target fish.

[0221] Optionally, the target recognition module further comprises an environment monitoring submodule and an adaptive adversarial sample generation network submodule, and the target recognition module is further configured to: acquire real-time environment data corresponding to the initial image sequence collected by the environment monitoring submodule; if the real-time environment data is greater than a preset threshold, control the adaptive adversarial sample generation network submodule to generate an adversarial sample corresponding to the real-time environment data; combine the initial sample corresponding to the target recognition result and the adversarial sample as training data, and train the scale perception backbone network submodule, the dual-stream spatiotemporal attention network submodule, the multi-scale feature passivation network submodule, and the multi-stage serial classification submodule based on the training data respectively, to obtain a training result; and dynamically update model parameters in the scale perception backbone network submodule, the dual-stream spatiotemporal attention network submodule, the multi-scale feature passivation network submodule, and the multi-stage serial classification submodule based on the training result.

[0222] Optionally, the target positioning module comprises a fluid topological visual SLAM sub-module, a deep tight-coupling feature enhancement network sub-module, and a double-state extended Kalman filter sub-module, wherein the fluid topological visual SLAM sub-module comprises a refraction distortion compensation module, a contour-enhanced feature point tracking module, and a dynamic continuous map construction module, and the second processing module 803 can be further configured to: perform distortion compensation processing on the target pixel coordinates and the target image sequence based on the refraction distortion compensation module to obtain a distortion-free image sequence; perform contour enhancement and feature point tracking processing on the distortion-free image sequence based on the contour-enhanced feature point tracking module to obtain a set of static feature points of the underwater environment in the distortion-free image sequence; and control the dynamic continuous map construction module to construct a target map based on the set of static feature points, and determine the carrier pose data according to the target map.

[0223] Optionally, the deep tight-coupling feature enhancement network sub-module comprises a three-dimensional observation vector construction module, a motion feature extraction module, and a target kinematic constraint module, and the double-state extended Kalman filter sub-module comprises an entity state predictor, a target trajectory predictor, and a confidence evaluation calibration module, and the third processing module 804 can be further configured to: control the three-dimensional observation vector construction module to perform fusion processing on the target pixel coordinates and the inertial measurement unit data to obtain a three-dimensional observation vector of the target fish in a coordinate system of the intelligent fishing system; extract relative motion features of the target fish and the intelligent fishing system and motion trend features of the intelligent fishing system from the three-dimensional observation vector based on the motion feature extraction module; perform kinematic constraint processing on the relative motion features and the motion trend features based on the target kinematic constraint module to obtain depth information of the target fish and the intelligent fishing system; control the entity state predictor to perform filtering processing on the carrier pose data and the inertial measurement unit data to obtain target carrier pose data, and convert the target pixel coordinates into three-dimensional coordinates through the target carrier pose data to obtain an initial positioning result; control the target trajectory predictor to perform filtering processing based on the initial positioning result, a historical position sequence, and target motion features to obtain an initial predicted trajectory result; and perform confidence evaluation calibration processing on the initial positioning result and the initial predicted trajectory result based on the confidence evaluation calibration module to obtain a target positioning result and a target predicted trajectory result.

[0224] In some embodiments, the intelligent fishing device 800 can further comprise a target decision module, which can be configured to: acquire target dynamic data of the target fish and visual quality evaluation data corresponding to the target image sequence; control the target decision module to perform deep learning processing based on the target dynamic data, the visual quality evaluation data, and the inertial measurement unit data to determine a first fishing strategy for capturing the target fish; and control the harpoon launching device to release a harpoon for fishing operation according to the first fishing strategy based on the target recognition result, the target positioning result, and the target predicted trajectory result.

[0225] Optionally, the target decision module comprises a spatio-temporal adaptive fusion network submodule and a decision-driven deep reinforcement learning submodule, wherein the decision-driven deep reinforcement learning submodule comprises a hierarchical state space module and a decision reinforcement policy generation network module, and the target decision module is further configured to: perform data fusion on the target dynamic data, the visual quality assessment data, and the inertial measurement unit data to construct multi-source input data; perform feature extraction on the multi-source input data from a time dimension and a space dimension based on the spatio-temporal adaptive fusion network submodule to determine first target motion features and first carrier disturbance features, and perform weighted processing on the first target motion features and the first carrier disturbance features to obtain a spatio-temporal fusion feature vector; perform hierarchical processing on the feature vector based on the hierarchical state space module to obtain a hierarchical processing result; and control the decision reinforcement policy generation network module to generate a target decision vector based on the hierarchical processing result and determine the first fishing strategy based on the target decision vector.

[0226] Optionally, the target decision module further comprises a multi-dimensional reward function submodule, and the decision-driven deep reinforcement learning submodule further comprises a decision reinforcement policy evaluation module, and the target decision module is further configured to: obtain historical fishing results of fishing operations performed according to the first fishing strategy; perform multi-dimensional analysis and processing on the historical fishing results based on the multi-dimensional reward function submodule to obtain a reward result; control the decision reinforcement policy evaluation module to evaluate the first fishing strategy based on the hierarchical processing result and the reward result to obtain an evaluation result; and if the evaluation result meets a preset condition, optimize the first fishing strategy and control the decision reinforcement policy generation network module to generate an updated first fishing strategy.

[0227] In some embodiments, the target decision module is further configured to: obtain target state data of the target fish, current carrier pose data of the intelligent fishing system, and environmental disturbance data, wherein the target state data comprises target three-dimensional coordinates corresponding to the target positioning result, a target predicted trajectory corresponding to the target predicted trajectory result, and a relative velocity vector of the target fish and the intelligent fishing system; control the target decision module to adjust the carrier pose of the intelligent fishing system based on the target state data, the current carrier pose data, and the environmental disturbance data to determine a second fishing strategy for capturing the target fish; and control the harpoon launching device to release the harpoon to perform fishing operations according to the first fishing strategy and / or the second fishing strategy based on the target recognition result, the target positioning result, and the target predicted trajectory result.

[0228] Optionally, the target decision module further comprises a double-branch spatiotemporal coding submodule, a pose solution output network submodule, and a turbine power distribution network submodule, wherein the double-branch spatiotemporal coding submodule comprises a target motion feature branch module, a carrier disturbance feature branch module, and an environmental attention fusion module. The target decision module can be further configured to: perform feature extraction on the target state data based on the target motion feature branch module to obtain second target motion features; perform feature extraction on the current carrier pose data and the environmental disturbance data based on the carrier disturbance feature branch module to obtain second carrier disturbance features; perform weighted fusion processing on the second target motion features and the second carrier disturbance features through the environmental attention fusion module to obtain a first fusion feature vector; perform processing on the first fusion feature vector through the pose solution output network submodule to obtain first angle control data corresponding to the intelligent fishing system; control the turbine power distribution network submodule to generate a first power distribution matrix based on the first angle control data; and perform power distribution on the turbines in the intelligent fishing system through the first power distribution matrix to obtain a second fishing strategy.

[0229] Optionally, the target decision module further comprises a feedback compensation network submodule. The target decision module can be further configured to: obtain first carrier pose data after the carrier pose adjustment of the intelligent fishing system through the second fishing strategy; perform analysis and processing on the first carrier pose data and the power distribution matrix based on the feedback compensation network submodule to obtain angle deviation compensation data; control the turbine power distribution network submodule to adjust the first angle control data based on the angle deviation compensation data to generate a second power distribution matrix; and perform power distribution on the turbines in the intelligent fishing system through the second power distribution matrix to obtain an updated second fishing strategy.

[0230] In some embodiments, the intelligent fishing device 800 can further comprise a target execution module, which can be configured to: control the target execution module to sequentially perform data fusion and precision optimization processing based on the target three-dimensional coordinates, the target predicted trajectory, and the inertial measurement unit data to generate a steering engine driving signal.

[0231] Correspondingly, the control module 805 can be further configured to: control the harpoon launching device to release the harpoon for fishing operations based on the steering engine driving signal.

[0232] Optionally, the target execution module comprises a self-calibration feature extraction network submodule, a double-branch precision control network submodule, and a multi-stage control instruction generation submodule, wherein the double-branch precision control network submodule comprises a timing precision control branch module, a disturbance suppression branch module, and a double-path collaborative fusion module, the multi-stage control instruction generation submodule comprises a microsecond-level trigger module, a super-precision angle control module, and a dynamic phase compensation module, and the target execution module can be further configured to: perform weighted fusion processing on the target three-dimensional coordinates, the target predicted trajectory, and the inertial measurement unit data based on the self-calibration feature extraction network submodule to obtain a second fusion feature vector; obtain historical rudder driving signals of a rudder in the intelligent fishing system and current environmental data underwater in a water area to be detected; perform deviation analysis processing on the historical rudder driving signals by the timing precision control branch module to obtain control deviation compensation data; perform weighted processing on the current environmental data by the disturbance suppression branch module to obtain environmental disturbance compensation data; control the double-path collaborative fusion module to sequentially perform compensation and fusion processing on the second fusion feature vector based on the control deviation compensation data and the environmental disturbance compensation data to obtain a target fusion feature vector; and perform precision optimization processing on the target fusion feature vector based on the microsecond-level trigger module, the super-precision angle control module, and the dynamic phase compensation module to obtain the rudder driving signals.

[0233] Optionally, the target execution module further comprises a closed-loop precision feedback network submodule, wherein the closed-loop precision feedback network submodule comprises an error compensation model and a high-precision detection encoder, and the target execution module can be further configured to: obtain actual control data of the rudder collected by the high-precision detection encoder; perform error comparison between the actual control data and target control data corresponding to the rudder driving signals to determine a control error result; control the error compensation model to perform weighted compensation processing on the second fusion feature vector according to the control error result to obtain a third fusion feature vector after compensation; and generate updated rudder driving signals based on the third fusion feature vector.

[0234] In some embodiments, the intelligent fishing device 800 can further comprise a target evolution module, which can be configured to: obtain an instantaneous image corresponding to a target fishing operation; perform analysis processing on the instantaneous image to obtain a fishing efficiency index, and optimize a target positioning result and a harpoon launching strategy of a harpoon launching device based on the fishing efficiency index to obtain an optimized control strategy; obtain an initial intelligent fishing edge model corresponding to the target fishing operation, and a shore-based large water-bottom world model; control the target evolution module to update and upgrade the initial intelligent fishing edge model based on the optimized control strategy and the shore-based large water-bottom world model to obtain a target intelligent fishing edge model.

[0235] Optionally, the target evolution module includes a multi-branch deduction evaluation network submodule, a fishing efficiency index fusion network submodule, and a double closed-loop collaborative optimization network submodule, wherein the double closed-loop collaborative optimization network submodule includes an identification positioning closed loop module and a decision promotion collaborative closed loop module, and the target evolution module can be further used for: analyzing and processing the instantaneous image based on the multi-branch deduction evaluation network submodule to obtain an evaluation result of the target fishing operation; performing weighted fusion processing on the evaluation result based on the fishing efficiency index fusion network submodule to obtain a fishing efficiency index; and performing positioning optimization on the target positioning result based on the fishing efficiency index and strategy optimization on the harpoon launch strategy through the double closed-loop collaborative optimization network submodule to obtain an optimized control strategy.

[0236] In some embodiments, the intelligent fishing device 800 can further include a safety control module, which can be used for: performing data synchronization and filtering processing on the inertial measurement unit data and the target image sequence through the data synchronization and preprocessing network submodule to obtain multi-degree-of-freedom data; if the multi-degree-of-freedom data is greater than a preset degree-of-freedom threshold, marking the intelligent fishing system as faulty through the safety control module and locking the intelligent fishing system.

[0237] Optionally, the safety control module further includes a time-frequency-space dynamic analysis network submodule, an attention fusion gate network submodule, a time sequence memory enhancement network submodule, and a dynamic stability decision network submodule, and the safety control module can be further used for: performing feature risk analysis on the multi-degree-of-freedom data from the time domain dimension, the frequency domain dimension, and the spatial dimension through the time-frequency-space dynamic analysis network submodule to obtain multi-dimensional risk features; performing weighted fusion processing on the multi-dimensional risk features based on the attention fusion gate network submodule to obtain an instantaneous risk feature vector; performing analysis processing on the instantaneous risk feature vector based on the time sequence memory enhancement network submodule to obtain a dynamic risk feature vector; performing stability analysis on the dynamic risk feature vector through the dynamic stability decision network submodule to obtain a stability probability result and a species risk distribution result, and determining whether to unlock or lock the intelligent fishing system based on the stability probability result and the species risk distribution result.

[0238] In some embodiments, the intelligent fishing device 800 can further include a posture control module, which can be used to: acquire target image and target inertial measurement unit data corresponding to the time when the harpoon launching device releases the harpoon, and perform fusion processing on the target image and the target inertial measurement unit data through a cross-modal data fusion network submodule to obtain fusion data; based on a dynamic feature fusion engine submodule, the fusion data is subjected to feature analysis and weighting processing respectively to obtain a comprehensive disturbance feature vector; the historical motion trajectory of the intelligent fishing system before the harpoon launching device releases the harpoon is acquired, and the historical motion trajectory and the comprehensive disturbance feature vector are integrated and analyzed through a motion trajectory integration network submodule to obtain a posture deviation amount; through an execution optimization network submodule, the real-time compensation thrust vector of the turbine in the intelligent fishing system is calculated based on the posture deviation amount, and the target turbine control parameter is obtained based on the real-time compensation thrust vector, so as to control the turbine through the target turbine control parameter.

[0239] It should be explained that the intelligent fishing device provided by the embodiments of the present application and the intelligent fishing method in the above embodiments belong to the same concept. Any method provided in the intelligent fishing method embodiments can be run on the intelligent fishing device. The specific implementation process is detailed in the intelligent fishing method embodiments, which will not be described here.

[0240] In the present embodiment, the intelligent fishing device 800 acquires a target image sequence of the underwater environment in the water area to be detected containing fish through the acquisition module 801; the first processing module 802 performs feature extraction and feature classification processing based on the target image sequence in sequence to determine the target recognition result of the target fish; the second processing module 803 determines the carrier pose data of the intelligent fishing system according to the target pixel coordinates corresponding to the target recognition result and the target image sequence, and acquires the historical position sequence and the target motion feature of the target fish, as well as the inertial measurement unit data of the intelligent fishing system; the third processing module 804 performs filtering processing based on the carrier pose data, the inertial measurement unit data, the historical position sequence, and the target motion feature to determine the target positioning result and the target prediction trajectory result of the target fish; based on the target recognition result, the target positioning result, and the target prediction trajectory result, the control module 805 controls the harpoon launching device to release the harpoon for fishing operation, so as to improve the accuracy of fishing.

[0241] Optionally, the present embodiment further provides a computer readable storage medium, which stores computer program codes, when the computer program codes run on a computer, the computer program codes make the computer execute the related method steps to realize the intelligent fishing method provided by the above embodiments.

[0242] The storage medium can include a read-only memory (ROM), a random access memory (RAM), a magnetic disk, an optical disk, or the like.

[0243] Due to the instructions stored in the storage medium, the steps of any of the intelligent fishing methods provided in the embodiments of the present application can be executed, thus achieving the beneficial effects of any of the intelligent fishing methods provided in the embodiments of the present application. Details are described in the foregoing embodiments, which will not be repeated here.

[0244] Correspondingly, the embodiments of the present application provide an intelligent fishing system. Please refer to Figure 9 to Figure 11 , Figure 9 is a first structural schematic diagram of the intelligent fishing system provided in the embodiments of the present application, Figure 10 is a second structural schematic diagram of the intelligent fishing system provided in the embodiments of the present application, Figure 11 is a structural schematic diagram of a harpoon launching device. The intelligent fishing system 900 is a waterborne body capable of underwater fishing, such as an underwater unmanned aerial vehicle. The intelligent fishing system 900 can include a harpoon launching device 901, a rudder 902, and a buckle device 903.

[0245] Specifically, the harpoon launching device 901 includes a harpoon 9011 and an elastic structure 9012. One end of the elastic structure 9012 is connected to the harpoon 9011, and the elastic structure 9012 releases elastic potential energy to release the harpoon 9011 from the harpoon launching device 901. The harpoon launching device 901 is a high-strength alloy harpoon launching device with fast response, high penetration, and waterproof and pressure-resistant. The harpoon launching device 901 stores elastic potential energy through the elastic structure 9012, and the elastic structure 9012 can be a pre-compressed high-elasticity modulus energy storage spring.

[0246] The output shaft of the rudder 902 is connected to the rotating shaft of the buckle device 903, and the rudder 902 is used to control the rotation of the buckle device 903.

[0247] One end of the buckle device 903 abuts against the other end of the elastic structure 9012, and the buckle device 903 controls the release or storage of elastic potential energy of the elastic structure 9012 through the rotation of the rotating shaft. The buckle device 903 is a rotary buckle device driven by the rudder 902, which is used to constrain the elastic structure 9012 from rebounding. The output shaft of the rudder 902 is connected to the rotating shaft of the buckle device 903, and the buckle device 903 is triggered to be unlocked instantaneously, so that the harpoon launching device 901 releases the harpoon 9011.

[0248] It should be noted that the steering engine 902 has the characteristics of high torque design, fast response optimization, and waterproof and pressure-resistant reinforcement. Specifically, based on the high torque design characteristic, the pre-tightening force of the elastic structure 9012 needs to be overcome, so the steering engine 902 selects a large reduction ratio gear set to ensure that the output torque can tightly clamp the buckle device 903; based on the fast response optimization characteristic, the internal motor of the steering engine 902 uses a high-speed model, which cooperates with a short pulse trigger signal to enable the output shaft to complete the unlocking rotation within microseconds; based on the waterproof and pressure-resistant reinforcement characteristic, the steering engine 902 works in an underwater environment, so the shell of the steering engine 902 needs to be waterproof and sealed, and high-strength alloy materials are selected to adapt to the high-pressure underwater environment, so that the internal components such as the motor and the gear set are not affected by water and pressure.

[0249] In addition, it should be noted that the intelligent fishing system 900 can also include an intelligent module and a relay. The relay can control the circuit of the intelligent fishing system 900 through signals such as current and voltage. The intelligent module can be a kind of controller, such as an edge processor. The intelligent module can control the steps in the above-mentioned intelligent fishing method embodiments, and the intelligent module can also control the modules in the intelligent fishing device to realize corresponding functions.

[0250] Correspondingly, the present application also provides an intelligent fishing system. Please refer to Figure 12 , Figure 12 is a third structure diagram of the intelligent fishing system provided by the present application. The intelligent fishing system 900 includes a processor 904 and a memory 905. The processor 904 is electrically connected with the memory 905.

[0251] The processor 904 is the control center of the intelligent fishing system 900, which connects all parts of the intelligent fishing system through various interfaces and lines, executes various functions of the intelligent fishing system and processes data by running or calling the computer programs stored in the memory 905 and calling the data stored in the memory 905, so as to monitor the whole intelligent fishing system.

[0252] The memory 905 can be used to store software programs and modules. The processor 904 executes various functions and data processing by running the computer programs and modules stored in the memory 905. The memory 905 can mainly include a program storage area and a data storage area. The program storage area can store an operating system, computer programs required by at least one function, etc.; the data storage area can store data created according to the use of the intelligent fishing system, etc.

[0253] In addition, the memory 905 can include a high-speed random access memory, and can also include a non-volatile memory, such as at least one magnetic disk storage device, a flash memory device, or other volatile solid-state memory device. Accordingly, the memory 905 can also include a memory controller to provide the processor 904 with access to the memory 905.

[0254] In the embodiment, the processor 904 in the intelligent fishing system 900 loads the instructions corresponding to the processes of one or more computer programs into the memory 905 according to the following steps, and runs the computer programs stored in the memory 905 by the processor 904, so as to realize various functions, as follows:

[0255] Obtain a target image sequence of an underwater environment in a water area containing fish to be detected;

[0256] Control the target recognition module to sequentially perform feature extraction and feature classification processing based on the target image sequence, and determine a target recognition result of the target fish;

[0257] Determine carrier pose data of the intelligent fishing system, and obtain a historical position sequence and a target motion feature of the target fish, and inertial measurement unit data of the intelligent fishing system, according to the target pixel coordinates corresponding to the target recognition result and the target image sequence;

[0258] Control the target positioning module to perform filtering processing based on the carrier pose data, the inertial measurement unit data, the historical position sequence, and the target motion feature, and determine a target positioning result and a target prediction trajectory result of the target fish;

[0259] Control the harpoon launching device to release the harpoon to perform fishing operations based on the target recognition result, the target positioning result, and the target prediction trajectory result.

[0260] In the embodiments provided in the present application, it should be understood that the disclosed devices and methods can be implemented in other ways. For example, the device embodiments described above are only schematic. For example, the division of the modules or units is only a logical function division, and actual implementation can have another division manner. For example, a plurality of units or components can be combined or integrated into another device, or some features can be ignored or not executed. In addition, the coupling or direct coupling or communication connection between the units or components shown or discussed can be indirect coupling or communication connection through some interfaces, devices or units, and can be electrical, mechanical or other forms.

[0261] For the intelligent fishing device of the embodiments of the present application, each function module can be integrated in one processing chip, or each module can exist physically alone, or two or more modules can be integrated in one module. The integrated module can be realized in the form of hardware or in the form of a software function module. When the integrated module is realized in the form of a software function module and sold or used as an independent product, it can also be stored in a computer readable storage medium.

[0262] The intelligent fishing method, device, system and storage medium provided by the embodiments of the present application are described in detail above. The principles and implementation manners of the present application are described by applying specific examples in this paper, and the above embodiments are only used to help understand the method of the present application and its core idea; meanwhile, for those skilled in the art, according to the idea of the present application, the specific implementation manner and application range will be changed, and the above description should not be understood as a limitation on the present application.

Claims

1. A method of smart fishing, characterized in that, The method is applied to an intelligent fishing system, and the intelligent fishing system at least comprises a target recognition module, a target positioning module, a target decision module and a harpoon launching device. A target image sequence of an underwater environment in a water area containing fish to be detected is acquired. The target recognition module is controlled to sequentially perform feature extraction and feature classification processing based on the target image sequence to determine a target recognition result of a target fish. Carrier pose data of the intelligent fishing system is determined according to target pixel coordinates corresponding to the target recognition result and the target image sequence, and historical position sequence and target motion characteristics of the target fish are acquired, and inertial measurement unit data of the intelligent fishing system is acquired. The target positioning module is controlled to perform filtering processing based on the carrier pose data, the inertial measurement unit data, the historical position sequence and the target motion characteristics to determine a target positioning result and a target prediction trajectory result of the target fish. Target dynamic data of the target fish and visual quality evaluation data corresponding to the target image sequence are acquired. The target decision module is controlled to perform deep learning processing based on the target dynamic data, the visual quality evaluation data and the inertial measurement unit data to determine a first fishing strategy for capturing the target fish. The harpoon launching device is controlled to release a harpoon for fishing operation according to the first fishing strategy based on the target recognition result, the target positioning result and the target prediction trajectory result.

2. The intelligent fishing method of claim 1, wherein, After the target positioning module is controlled to perform filtering processing based on the carrier pose data, the inertial measurement unit data, the historical position sequence and the target motion characteristics to determine the target positioning result and the target prediction trajectory result of the target fish, the method further comprises: Target state data of the target fish, current carrier pose data of the intelligent fishing system and environmental interference data are acquired, wherein the target state data comprises target three-dimensional coordinates corresponding to the target positioning result, a target prediction trajectory corresponding to the target prediction trajectory result and a relative speed vector of the target fish and the intelligent fishing system. The target decision module is controlled to adjust the carrier pose of the intelligent fishing system based on the target state data, the current carrier pose data and the environmental interference data to determine a second fishing strategy for capturing the target fish. The harpoon launching device is controlled to release a harpoon for fishing operation according to the first fishing strategy and / or the second fishing strategy based on the target recognition result, the target positioning result and the target prediction trajectory result.

3. The intelligent fishing method of claim 1, wherein, The intelligent fishing system further comprises a target execution module, and before the harpoon launching device is controlled to release a harpoon for fishing operation, the method further comprises: The target execution module is controlled to sequentially perform data fusion and precision optimization processing based on target three-dimensional coordinates, the target prediction trajectory and the inertial measurement unit data to generate a rudder driving signal. The control of the harpoon launching device to release a harpoon for fishing operation comprises: Based on the rudder driving signal, the fish spear launching device is controlled to release a fish spear for fishing operation.

4. The method of smart fishing according to any one of claims 1 to 3, characterized in that, The intelligent fishing system further comprises a target evolution module, and after the fish spear launching device is controlled to release a fish spear for fishing, the method further comprises: acquiring an instantaneous image corresponding to a target fishing operation; analyzing and processing the instantaneous image to obtain a fishing efficiency index, and optimizing the target positioning result and the fish spear launching strategy of the fish spear launching device based on the fishing efficiency index to obtain an optimized control strategy; acquiring an initial intelligent fishing edge model corresponding to the target fishing operation, and a large shore-based water-bottom world model; based on the optimized control strategy and the large shore-based water-bottom world model, the target evolution module is controlled to update and upgrade the initial intelligent fishing edge model to obtain a target intelligent fishing edge model.

5. The intelligent fishing method of claim 1, wherein, The target recognition module comprises an underwater optical compensation algorithm submodule, and the target image sequence of the underwater environment in the water area to be detected containing fish is acquired, comprising: acquiring an initial image sequence of the underwater environment in the water area to be detected; based on the optical compensation algorithm submodule, the initial image sequence is subjected to despeckling processing to eliminate turbid medium scattering noise in the initial image sequence to obtain the target image sequence.

6. The method of claim 5, wherein, The target recognition module further comprises a scale pattern perception backbone network submodule, a double-flow spatio-temporal attention network submodule, a multi-scale feature passivation network submodule, and a multi-level serial classification submodule, wherein the double-flow spatio-temporal attention network submodule comprises a spatial feature flow attention module, a temporal feature flow attention module, and a double-flow feature integration module, the target recognition module is controlled to sequentially perform feature extraction and feature classification processing based on the target image sequence to determine the target recognition result of the target fish, comprising: based on the scale pattern perception backbone network submodule, the target image sequence is subjected to feature extraction from spatial and temporal dimensions to obtain a spatio-temporal fusion feature map of fish; the spatial feature flow attention module is controlled to perform weighting processing on the spatio-temporal fusion feature map in the spatial dimension to obtain target morphological features of fish; the temporal feature flow attention module is controlled to obtain multiple motion features of the spatio-temporal fusion feature map in the temporal dimension to obtain target motion trajectory features of fish; based on the double-flow feature integration module, cross-dimension feature integration is performed on the target morphological features and the target motion trajectory features to obtain an initial feature vector of fish; the multi-scale feature passivation network submodule is controlled to sequentially perform multi-scale feature extraction and feature passivation processing on the initial feature vector to compress background feature interference to obtain a target feature vector; based on the multi-level serial classification submodule, the target feature vector is subjected to multi-level feature classification to determine the target recognition result of the target fish.

7. The method of claim 6, wherein, The target recognition module further comprises an environment monitoring submodule and an adaptive adversarial sample generation network submodule, and after the initial image sequence of the underwater environment in the water area to be detected is acquired, the method further comprises: acquiring real-time environmental data corresponding to the initial image sequence collected by the environment monitoring submodule; If the real-time environment data is greater than a preset threshold, the adaptive adversarial sample generation network submodule is controlled to generate an adversarial sample corresponding to the real-time environment data; The initial sample corresponding to the target recognition result is combined with the adversarial sample as training data, and the training data is used to train the scale pattern perception backbone network submodule, the double-flow space-time attention network submodule, the multi-scale feature passivation network submodule, and the multi-stage serial classification submodule respectively to obtain training results; Based on the training results, the model parameters in the scale pattern perception backbone network submodule, the double-flow space-time attention network submodule, the multi-scale feature passivation network submodule, and the multi-stage serial classification submodule are dynamically updated.

8. The intelligent fishing method of claim 1, wherein, The target positioning module includes a fluid topology visual SLAM submodule, a deep tight coupling feature enhancement network submodule, and a double-state extended Kalman filter submodule. The fluid topology visual SLAM submodule includes a refraction distortion compensation module, a contour-enhanced feature point tracking module, and a dynamic continuous map construction module. The carrier pose data of the intelligent fishing system is determined based on the target pixel coordinates corresponding to the target recognition result and the target image sequence, which includes: The refraction distortion compensation module is used to perform distortion compensation processing on the target pixel coordinates and the target image sequence to obtain a distortion-free image sequence; The contour-enhanced feature point tracking module is used to perform contour enhancement and feature point tracking processing on the distortion-free image sequence to obtain a set of static feature points of the underwater environment in the distortion-free image sequence; The dynamic continuous map construction module is controlled to construct a target map based on the set of static feature points, and to determine the carrier pose data based on the target map.

9. The intelligent fishing method of claim 8, wherein, The deep tight coupling feature enhancement network submodule includes a three-dimensional observation vector construction module, a motion feature extraction module, and a target kinematics constraint module. The double-state extended Kalman filter submodule includes an entity state predictor, a target trajectory predictor, and a confidence evaluation calibration module. The target positioning module is controlled to perform filtering processing based on the carrier pose data, the inertial measurement unit data, the historical position sequence, and the target motion feature to determine the target positioning result and the target prediction trajectory result of the target fish, which includes: The three-dimensional observation vector construction module is controlled to fuse the target pixel coordinates and the inertial measurement unit data to obtain a three-dimensional observation vector of the target fish in the coordinate system of the intelligent fishing system; The motion feature extraction module is used to extract the relative motion feature of the target fish and the intelligent fishing system and the motion trend feature of the intelligent fishing system from the three-dimensional observation vector; The target kinematics constraint module is used to perform kinematics constraint processing on the relative motion feature and the motion trend feature to obtain the depth information of the target fish and the intelligent fishing system; The body state predictor is controlled to filter the carrier pose data and the inertial measurement unit data to obtain target carrier pose data, and the target pixel coordinates are converted into three-dimensional coordinates through the target carrier pose data to obtain an initial positioning result; The target trajectory predictor is controlled to filter the initial positioning result, the historical position sequence, and the target motion feature to obtain an initial predicted trajectory result; The confidence evaluation calibration module is used to perform confidence evaluation calibration processing on the initial positioning result and the initial predicted trajectory result to obtain the target positioning result and the target predicted trajectory result.

10. The intelligent fishing method of claim 1, wherein, The target decision module includes a spatio-temporal adaptive fusion network submodule and a decision-driven deep reinforcement learning submodule. The decision-driven deep reinforcement learning submodule includes a hierarchical state space module and a decision reinforcement policy generation network module. The target decision module is controlled to perform deep learning processing based on the target dynamic data, the visual quality evaluation data, and the inertial measurement unit data to determine a first fishing strategy for capturing the target fish, including: The target dynamic data, the visual quality evaluation data, and the inertial measurement unit data are fused to construct multi-source input data; The spatio-temporal adaptive fusion network submodule is used to extract features from the multi-source input data in the time dimension and the space dimension to determine first target motion features and first carrier disturbance features, and the first target motion features and the first carrier disturbance features are weighted to obtain a spatio-temporal fusion feature vector; The hierarchical state space module is used to perform hierarchical processing on the feature vector to obtain a hierarchical processing result; The decision reinforcement policy generation network module is controlled to generate a target decision vector based on the hierarchical processing result, and the first fishing strategy is determined based on the target decision vector.

11. The intelligent fishing method of claim 10, wherein, The target decision module further includes a multi-dimensional reward function submodule, and the decision-driven deep reinforcement learning submodule further includes a decision reinforcement policy evaluation module. After the fish spear launching device is controlled to release a fish spear to perform fishing operations according to the first fishing strategy, the method further includes: A historical fishing result of performing fishing operations according to the first fishing strategy is obtained; The multi-dimensional reward function submodule is used to perform multi-dimensional analysis processing on the historical fishing result to obtain a reward result; The decision reinforcement policy evaluation module is controlled to evaluate the first fishing strategy based on the hierarchical processing result and the reward result to obtain an evaluation result; If the evaluation result meets a preset condition, the first fishing strategy is optimized, and the decision reinforcement policy generation network module is controlled to generate an updated first fishing strategy.

12. The intelligent fishing method of claim 2, wherein, The target decision module further comprises a double-branch space-time coding submodule, a pose solution output network submodule, and a turbine power distribution network submodule, wherein the double-branch space-time coding submodule comprises a target motion feature branch module, a carrier disturbance feature branch module, and an environmental attention fusion module; the control of the target decision module adjusts the carrier pose of the intelligent fishing system based on the target state data, the current carrier pose data, and the environmental disturbance data, determines a second fishing strategy for capturing the target fish, comprising: extracting features from the target state data based on the target motion feature branch module to obtain second target motion features; extracting features from the current carrier pose data and the environmental disturbance data based on the carrier disturbance feature branch module to obtain second carrier disturbance features; performing weighted fusion processing on the second target motion features and the second carrier disturbance features through the environmental attention fusion module to obtain a first fusion feature vector; processing the first fusion feature vector through the pose solution output network submodule to obtain first angle control data corresponding to the intelligent fishing system; controlling the turbine power distribution network submodule to generate a first power distribution matrix based on the first angle control data; performing power distribution on the turbines in the intelligent fishing system through the first power distribution matrix to obtain the second fishing strategy.

13. The method of claim 12, wherein, The target decision module further comprises a feedback compensation network submodule, and after the power distribution matrix is used to perform power distribution on the turbines in the intelligent fishing system to obtain the second fishing strategy, the method further comprises: obtaining first carrier pose data after the carrier pose adjustment of the intelligent fishing system through the second fishing strategy; analyzing and processing the first carrier pose data and the power distribution matrix based on the feedback compensation network submodule to obtain angle deviation compensation data; controlling the turbine power distribution network submodule to adjust the first angle control data based on the angle deviation compensation data to generate a second power distribution matrix; performing power distribution on the turbines in the intelligent fishing system through the second power distribution matrix to obtain an updated second fishing strategy.

14. The intelligent fishing method of claim 3, wherein, The target execution module comprises a self-calibration feature extraction network submodule, a double-branch precision control network submodule, and a multi-stage control instruction generation submodule, wherein the double-branch precision control network submodule comprises a timing precision control branch module, a disturbance suppression branch module, and a double-path collaborative fusion module, the multi-stage control instruction generation submodule comprises a microsecond-level trigger module, a super-precision angle control module, and a dynamic phase compensation module, and the control of the target execution module sequentially performs data fusion and precision optimization processing based on the target three-dimensional coordinates, the target predicted trajectory, and the inertial measurement unit data to generate a rudder drive signal, comprising: performing weighted fusion processing on the target three-dimensional coordinates, the target predicted trajectory, and the inertial measurement unit data based on the self-calibration feature extraction network submodule to obtain a second fusion feature vector; Acquire the historical rudder drive signal of the rudder in the intelligent fishing system and the current underwater environment data in the water area to be detected; Perform deviation analysis processing on the historical rudder drive signal through the timing precision control branch module to obtain control deviation compensation data; Perform weighting processing on the current environment data through the disturbance suppression branch module to obtain environment interference compensation data; Control the double-path collaborative fusion module to sequentially perform compensation and fusion processing on the second fusion feature vector based on the control deviation compensation data and the environment interference compensation data to obtain a target fusion feature vector; Perform precision optimization processing on the target fusion feature vector based on the microsecond-level trigger module, the ultra-precision angle control module, and the dynamic phase compensation module to obtain the rudder drive signal.

15. The method of claim 14, wherein, The target execution module further includes a closed-loop precision feedback network submodule, wherein the closed-loop precision feedback network submodule includes an error compensation model and a high-precision detection encoder. After performing precision optimization processing on the target fusion feature vector based on the microsecond-level trigger module, the ultra-precision angle control module, and the dynamic phase compensation module to obtain the rudder drive signal, the method further includes: Acquire the actual control data of the rudder collected by the high-precision detection encoder; Perform error comparison between the actual control data and the target control data corresponding to the rudder drive signal to determine a control error result; According to the control error result, control the error compensation model to perform weighted compensation processing on the second fusion feature vector to obtain a third fusion feature vector after compensation; Generate an updated rudder drive signal based on the third fusion feature vector.

16. The intelligent fishing method of claim 4, wherein, The target evolution module includes a multi-branch deductive evaluation network submodule, a fishing efficiency index fusion network submodule, and a double-closed-loop collaborative optimization network submodule. The double-closed-loop collaborative optimization network submodule includes an identification and positioning closed-loop module and a decision and propulsion collaborative closed-loop module. The method includes: Perform analysis processing on the instantaneous image based on the multi-branch deductive evaluation network submodule to obtain an evaluation result of the target fishing operation; Perform weighted fusion processing on the evaluation result based on the fishing efficiency index fusion network submodule to obtain the fishing efficiency index; Perform positioning optimization on the target positioning result and strategy optimization on the harpoon launching strategy of the harpoon launching device based on the fishing efficiency index through the double-closed-loop collaborative optimization network submodule to obtain the optimized control strategy.

17. The method of smart fishing according to any one of claims 1 to 3, wherein, The intelligent fishing system further includes a safety control module. The safety control module includes a data synchronization and preprocessing network submodule. Before controlling the harpoon launching device to release a harpoon for fishing operation, the method further includes: The inertial measurement unit data and the target image sequence are subjected to data synchronization and filtering processing by the data synchronization and preprocessing network submodule, to obtain multi-degree-of-freedom data; If the multi-degree-of-freedom data is greater than a preset degree-of-freedom threshold, the intelligent fishing system is marked as faulty and locked by the safety control module.

18. The method of claim 17, wherein, The safety control module further comprises a time-frequency-space dynamic analysis network submodule, an attention fusion gate network submodule, a time-series memory enhancement network submodule, and a dynamic stability decision network submodule. After the intelligent fishing system is locked, the method further comprises: The multi-degree-of-freedom data is subjected to feature risk analysis from the time domain dimension, the frequency domain dimension, and the spatial dimension by the time-frequency-space dynamic analysis network submodule, to obtain multi-dimensional risk features; The multi-dimensional risk features are subjected to weighted fusion processing based on the attention fusion gate network submodule, to obtain an instantaneous risk feature vector; The instantaneous risk feature vector is subjected to analysis processing based on the time-series memory enhancement network submodule, to obtain a dynamic risk feature vector; The dynamic risk feature vector is subjected to stability analysis by the dynamic stability decision network submodule, to obtain a stability probability result and a species risk distribution result, and the intelligent fishing system is determined to be unlocked or locked based on the stability probability result and the species risk distribution result.

19. The method of smart fishing according to any one of claims 1 to 3, wherein, The intelligent fishing system further comprises a posture control module, which comprises a cross-modal data fusion network submodule, a dynamic feature fusion engine submodule, a motion trajectory integration network submodule, and an execution optimization network submodule. The method further comprises: The target image and the target inertial measurement unit data corresponding to the time when the harpoon launching device releases the harpoon are obtained, and the target image and the target inertial measurement unit data are subjected to fusion processing by the cross-modal data fusion network submodule, to obtain fusion data; The fusion data is subjected to feature analysis and weighted processing based on the dynamic feature fusion engine submodule, to obtain a comprehensive disturbance feature vector; The historical motion trajectory of the intelligent fishing system before the harpoon launching device releases the harpoon is obtained, and the historical motion trajectory and the comprehensive disturbance feature vector are subjected to integration analysis by the motion trajectory integration network submodule, to obtain a posture deviation amount; The real-time compensation thrust vector of the turbine in the intelligent fishing system is calculated based on the posture deviation amount by the execution optimization network submodule, and target turbine control parameters are obtained based on the real-time compensation thrust vector, so that the turbine is controlled by the target turbine control parameters.

20. An intelligent fishing device, characterized by The device is applied to an intelligent fishing system, and the intelligent fishing system at least comprises a target identification module, a target positioning module, and a harpoon launching device. The device comprises: An acquisition module is configured to acquire a target image sequence of an underwater environment in a water area containing fish to be detected; A first processing module is configured to control the target identification module to sequentially perform feature extraction and feature classification processing based on the target image sequence, to determine a target identification result of a target fish; The second processing module is configured to determine carrier pose data of the intelligent fishing system according to target pixel coordinates corresponding to the target recognition result and the target image sequence, and to obtain a historical position sequence and target motion characteristics of the target fish, and inertial measurement unit data of the intelligent fishing system. The third processing module is configured to control the target positioning module to perform filtering processing based on the carrier pose data, the inertial measurement unit data, the historical position sequence, and the target motion characteristics, to determine a target positioning result and a target prediction trajectory result of the target fish. The control module is configured to obtain target dynamic data of the target fish and visual quality evaluation data corresponding to the target image sequence, to control the target decision module to perform deep learning processing based on the target dynamic data, the visual quality evaluation data, and the inertial measurement unit data, to determine a first fishing strategy for capturing the target fish, and to control the harpoon launching device to release a harpoon for fishing operation according to the first fishing strategy based on the target recognition result, the target positioning result, and the target prediction trajectory result.

21. An intelligent fishing system characterized by, The intelligent fishing system is configured to implement the intelligent fishing method according to any one of claims 1 to 19, and the intelligent fishing system comprises: a harpoon launching device, a steering engine, and a buckle device, wherein the harpoon launching device comprises a harpoon and an elastic structure, one end of the elastic structure is connected with the harpoon, and the elastic structure releases elastic potential energy to enable the harpoon launching device to release the harpoon; an output shaft of the steering engine is connected with a rotating shaft of the buckle device, and the steering engine is configured to control the buckle device to rotate; one end of the buckle device is in abutment with the other end of the elastic structure, and the buckle device controls the elastic structure to release or store elastic potential energy through rotation of the rotating shaft.

22. An intelligent fishing system characterized by, The intelligent fishing system comprises: a memory configured to store executable program codes; a processor configured to call and run the executable program codes from the memory, so that the intelligent fishing system performs the intelligent fishing method according to any one of claims 1 to 19.

23. A computer-readable storage medium, characterized in that, The computer readable storage medium stores a computer program, when the computer program is executed, the intelligent fishing method according to any one of claims 1 to 19 is implemented. The computer readable storage medium stores a computer program, when the computer program is executed, the intelligent fishing method according to any one of claims 1 to 19 is implemented.

Images