Image recognition based vehicle obstacle recognition method and device

Abstract

This application discloses a method and apparatus for obstacle recognition in submersibles based on image recognition, relating to the field of submersible technology. The method includes: constructing an obstacle recognition model and an obstacle avoidance model on a central server and deploying them to all submersibles; preprocessing multi-dimensional sensor data collected by the submersibles and inputting the preprocessed multi-dimensional sensor data into the obstacle recognition model for obstacle recognition, obtaining obstacle recognition results; inputting the obstacle recognition results into the obstacle avoidance model for obstacle avoidance decision generation, obtaining obstacle avoidance decisions, and generating corresponding obstacle avoidance commands; performing a safety assessment on the obstacle avoidance commands, and if the assessment passes, converting the obstacle avoidance commands into submersible control signals to drive the submersible to perform obstacle avoidance. This method solves the problems of insufficient recognition accuracy, poor model adaptability, and low intelligence in existing technologies.

Description

Technical Field

[0001] This invention relates to the field of submersible technology, and in particular to a method and apparatus for obstacle recognition in submersibles based on image recognition. Background Technology

[0002] With the increasing demand for marine resource development and marine tourism, submersibles are playing an increasingly important role in underwater operations. However, the underwater environment is complex and ever-changing, containing various static (such as reefs and shipwrecks) and dynamic (such as fish and other vehicles) obstacles, posing a severe challenge to the autonomous navigation and safe operation of submersibles. Therefore, achieving high-precision and high-reliability obstacle recognition and real-time, intelligent autonomous avoidance is a key technology for ensuring the success of submersible missions and their own safety.

[0003] Existing obstacle recognition technologies for submersibles have the following main drawbacks:

[0004] 1) Insufficient recognition accuracy: Traditional methods often rely on a single sensor, such as sonar alone. While sonar has a long range and strong penetration, its low resolution and blurry images make it difficult to identify small or complex-shaped obstacles. Optical cameras can provide high-resolution texture information, but their range is drastically reduced in turbid water and they are greatly affected by lighting conditions. The limitations of a single sensor lead to a high rate of missed and false detections of obstacles in complex and variable underwater environments.

[0005] 2) Poor model adaptability: Some existing technologies use traditional image processing algorithms (such as edge detection and threshold segmentation) for obstacle recognition. These algorithms have poor robustness and are difficult to cope with interference such as uneven underwater lighting, water scattering, and cluttered backgrounds. Although there are deep learning-based image recognition methods applied underwater, most of them only process single optical images and fail to effectively integrate the advantages of heterogeneous sensors such as sonar, resulting in a significant performance degradation of the model in extreme environments.

[0006] 3) Low intelligence: Traditional obstacle avoidance methods are mostly based on artificial rules (such as potential field methods, artificial potential field methods, and fuzzy logic control). These methods have complex rule designs, rely on expert experience, and are difficult to cover all possible underwater scenarios. For example, artificial potential field methods are prone to getting stuck in local minima when in narrow passages or facing multiple obstacles, leading to planning failure. These methods lack learning and adaptive capabilities and cannot dynamically optimize their behavior according to environmental changes. Summary of the Invention

[0007] This invention provides a method and apparatus for underwater obstacle recognition based on image recognition, which solves the problems of insufficient recognition accuracy, poor model adaptability and low intelligence in the existing technology.

[0008] In a first aspect, embodiments of the present invention provide a method for underwater vehicle obstacle recognition based on image recognition, the method comprising:

[0009] On the central server, an obstacle recognition model and an obstacle avoidance model are built and deployed to all submersibles.

[0010] The multidimensional sensor data collected by the submersible is preprocessed, and the preprocessed multidimensional sensor data is input into the obstacle recognition model to perform obstacle recognition and obtain the obstacle recognition result.

[0011] The obstacle recognition results are input into the obstacle avoidance model to generate obstacle avoidance decisions, and corresponding obstacle avoidance instructions are generated.

[0012] A safety assessment is performed on the obstacle avoidance command. If the assessment is successful, the obstacle avoidance command is converted into a submersible control signal to drive the submersible to avoid the obstacle.

[0013] The technical solution provided in this application has at least the following beneficial effects:

[0014] By monitoring multi-dimensional sensor data, combining the detail advantages of optical images with the long-range, all-weather advantages of sonar images, the limitations of single sensors are overcome, significantly improving obstacle recognition accuracy and recall under various water quality and lighting conditions. A reinforcement learning-based obstacle avoidance model replaces traditional manual rules, enabling the submersible to autonomously learn complex avoidance strategies through interaction with the environment, enhancing its intelligence. Its decision-making is no longer a rigid "if-then" logic, but rather generates smooth, efficient, and human-driver-like intelligent avoidance behavior based on the distribution of multiple obstacles, its own state, and mission objectives. An obstacle avoidance system constructed using image recognition algorithms is employed. Unlike traditional image processing methods, deep learning models can automatically learn and extract deep, abstract features of obstacles from massive amounts of data, rather than relying on simple, manually designed features (such as edges and colors). This allows the model to effectively distinguish complex underwater backgrounds (such as seaweed, rocks, and shadows) from real obstacles. It is also highly robust to interference from changes in lighting, water scattering, and partial occlusion, thereby raising the recognition accuracy to a new level. Furthermore, it introduces an independent safety assessment process to perform secondary verification of decisions, effectively preventing collisions caused by model misjudgments or extreme situations, and providing double protection for the safe operation of submersibles.

[0015] In one alternative implementation, an obstacle recognition model and an obstacle avoidance model are built on a central server and deployed to all submersibles, including:

[0016] On the central server, an initial obstacle recognition model is built using image recognition algorithms, an initial obstacle avoidance model is built using intelligent decision-making algorithms, and the initial obstacle recognition model and the initial obstacle avoidance model are deployed to submersibles in different scenarios.

[0017] In each scene of the submersible, based on the collected first scene dataset, the initial obstacle recognition model is trained for the scene, and the first optimization algorithm is introduced to optimize the hyperparameters of the obstacle recognition model during the scene training process, so as to obtain the scene obstacle recognition model and several scene obstacle recognition results.

[0018] Based on the obstacle recognition results of several scenarios, the corresponding submersible status of the scenarios, and the target information of the scenarios, a second scenario dataset is constructed. Based on the second scenario dataset, the initial obstacle avoidance model is trained on the scenarios. A second optimization algorithm is introduced to optimize the action selection mechanism of the obstacle avoidance model, thereby obtaining the obstacle avoidance model of the scenarios.

[0019] Extract the model update amount and scene data amount of the obstacle recognition model and the obstacle avoidance model of each submersible scene. Based on the update amount of all models and the scene data amount, update the initial obstacle recognition model and the initial obstacle avoidance model of the central server to obtain the global obstacle recognition model and the global obstacle avoidance model.

[0020] Deploy the global obstacle recognition model and the global obstacle avoidance model to all submersibles.

[0021] In one alternative implementation, the obstacle recognition model is constructed based on the EfficientNet-AFPN algorithm, and the obstacle recognition model includes a shared backbone network constructed based on the EfficientNet algorithm, a neck network constructed based on the AFPN algorithm, and a multi-task unified detection head. The multi-task unified detection head includes parallel type detection branches, position detection branches, size detection branches, and relative velocity detection branches, and the first optimization algorithm is the PRGO-PSO algorithm.

[0022] The obstacle avoidance model is built based on the PPO algorithm and includes an agent and an experience replay pool. The second optimization algorithm is the ACO algorithm.

[0023] In one optional implementation, for each scene in the submersible, based on the collected first scene dataset, an initial obstacle recognition model is trained for the scene, and a first optimization algorithm is introduced to optimize the hyperparameters of the obstacle recognition model during scene training, resulting in a scene obstacle recognition model and several scene obstacle recognition results, including:

[0024] In each scene of the submersible, the initial obstacle recognition model is trained on the scene based on the collected first scene dataset with the first prediction label, and the PRGO-PSO algorithm is introduced during the scene training process.

[0025] The hyperparameters of the obstacle recognition model are encoded into individual vectors of the PRGO algorithm, and the fitness function, PRGO population parameters, and maximum number of iterations are set.

[0026] Based on the PRGO population parameters, the initial PRGO population is obtained by initializing using the Tent chaotic mapping sequence; each PRGO individual in the PRGO population corresponds to an alternative hyperparameter.

[0027] Based on the first scenario dataset, the fitness function is used to obtain the fitness value of each initial PRGO individual, and the initial PRGO individual with the best fitness value is taken as the optimal solution.

[0028] Based on the fiber root growth behavior and lateral root growth behavior of the PRGO algorithm, the initial PRGO population is iteratively updated to obtain the updated first PRGO population.

[0029] Based on the improved primary root growth behavior of the PSO mechanism of the PRGO algorithm, the updated first PRGO population is iteratively updated to obtain the updated second PRGO population.

[0030] Based on the first scenario dataset, the fitness function is used to obtain the fitness value of each updated second PRGO individual, and the updated second PRGO individual with the best fitness value is updated as the optimal solution;

[0031] When the number of iterations reaches the maximum number of iterations or the fitness value of the optimal solution meets the requirements, the iterative update of the PRGO population is terminated, and the optimal solution of the current iteration is output.

[0032] The individual vectors of the PRGO individuals corresponding to the optimal solution are decoded to obtain the optimized hyperparameters, and the optimized obstacle recognition model is obtained based on the optimized hyperparameters.

[0033] The first scene dataset is input into the optimized obstacle recognition model for fine-tuning, resulting in the obstacle recognition model for the scene and obstacle recognition results for several scenes.

[0034] In one optional implementation, a second scene dataset is constructed based on obstacle recognition results from several scenes, the corresponding submersible state in the scene, and scene target information. The initial obstacle avoidance model is then trained on this second scene dataset. A second optimization algorithm is introduced to optimize the action selection mechanism of the obstacle avoidance model, resulting in a scene-specific obstacle avoidance model, including:

[0035] The obstacle recognition results for each scene, the corresponding submersible status, and the scene target information are associated, and a second prediction label is added to obtain the second scene dataset.

[0036] Based on the second scenario dataset, the initial obstacle avoidance model was trained in the scenario, and the pheromone mechanism of the ACO algorithm was introduced during the scenario training process.

[0037] A virtual pheromone map and pheromone guide are set up for the initial obstacle avoidance model, and a dynamic update rule is set for the virtual pheromone map based on the reward mechanism of the obstacle avoidance model to obtain an optimized obstacle avoidance model.

[0038] The second scenario dataset is input into the optimized obstacle avoidance model for fine-tuning to obtain the obstacle avoidance model for the scenario.

[0039] In one alternative implementation, the multidimensional sensor data includes optical image data, sonar scan data, and the submersible's own pose data;

[0040] Obstacle identification results include the type, location, size, and relative speed of the obstacle.

[0041] In one optional implementation, the multidimensional sensor data collected by the submersible is preprocessed, and the preprocessed multidimensional sensor data is input into an obstacle recognition model for obstacle recognition, resulting in obstacle recognition results, including:

[0042] Using the multi-source sensors of the submersible, multi-dimensional sensor data is collected, and the multi-dimensional sensor data is spatiotemporally synchronized, denoised, enhanced, and spatially aligned to obtain pre-processed multi-dimensional sensor data, which is then input into the obstacle recognition model.

[0043] Using a shared backbone network of an obstacle recognition model, multi-scale features are extracted from preprocessed multi-dimensional sensor data.

[0044] The neck network of the obstacle recognition model is used to perform weighted fusion of multi-scale features to obtain fused features;

[0045] Based on the fusion features, a multi-task unified detection head is used to generate obstacle recognition results.

[0046] In one optional implementation, the obstacle recognition result is input into the obstacle avoidance model to generate an obstacle avoidance decision, thereby obtaining the obstacle avoidance decision and generating corresponding obstacle avoidance instructions, including:

[0047] Collect the submersible status and target information corresponding to the obstacle recognition results, and input the obstacle recognition results, submersible status and target information into the obstacle avoidance model;

[0048] Based on the submersible's mission planning map, initialize the virtual pheromone map of the obstacle avoidance model;

[0049] The obstacle recognition results, the submersible status, and the scene target information are written into the state space of the agent. Several historical experiences are randomly extracted from the experience replay pool, and the action set of the historical experiences is written into the action space of the agent.

[0050] Based on the state space, the agent's policy network is used to generate an initial probability distribution in the action space, and the initial probability distribution is discretized to generate the corresponding candidate action set.

[0051] Using a pheromone guide, predict the future location of each candidate action in the candidate action set, and query the pheromone concentration at the future location on a virtual pheromone map;

[0052] Based on the pheromone concentration, calculate the pheromone bias weight for each candidate action in each candidate action set, and generate the final selection probability of the candidate action set based on the pheromone bias weight.

[0053] The final selection probability is transformed into a final probability distribution, and based on the final probability distribution, the action to be executed, i.e. the obstacle avoidance decision, is selected and the corresponding obstacle avoidance instruction is generated.

[0054] Based on the reward function, the agent generates reward values ​​for the actions performed, and updates the virtual pheromone map according to the reward values ​​based on the dynamic update rules, thus obtaining an updated virtual pheromone map.

[0055] The real-time experience generated by obstacle avoidance decision-making is stored in the experience replay pool, and the obstacle avoidance model is continuously trained based on the real-time experience and some historical experience to obtain an updated obstacle avoidance model.

[0056] In one optional implementation, a safety assessment is performed on the obstacle avoidance command. If the assessment passes, the obstacle avoidance command is converted into a submersible control signal to drive the submersible to perform the avoidance maneuver, including:

[0057] The obstacle avoidance command is checked for safety boundaries. If the safety boundary check passes, the process proceeds to the next step; otherwise, a safety alarm is issued and the obstacle avoidance command is blocked.

[0058] The system performs dynamic trajectory prediction and collision risk assessment on obstacle avoidance commands. If the obtained dynamic trajectory has no collision risk, it proceeds to the next step; otherwise, it issues a safety alarm and blocks the obstacle avoidance command.

[0059] Based on the historical obstacle avoidance commands from the previous moment, the current obstacle avoidance commands are processed for motion smoothing and converted into corresponding submersible control signals to drive the submersible to perform the avoidance maneuver.

[0060] Secondly, embodiments of the present invention provide a submersible obstacle recognition device based on image recognition, used to implement a submersible obstacle recognition method, the device comprising:

[0061] The model deployment unit is used to build obstacle recognition models and obstacle avoidance models on the central server, and to deploy the obstacle recognition models and obstacle avoidance models to all submersibles;

[0062] The obstacle recognition unit is used to preprocess the multi-dimensional sensor data collected by the submersible, and input the preprocessed multi-dimensional sensor data into the obstacle recognition model to perform obstacle recognition and obtain the obstacle recognition result.

[0063] The obstacle avoidance unit is used to input the obstacle recognition results into the obstacle avoidance model, generate obstacle avoidance decisions, obtain obstacle avoidance decisions, and generate corresponding obstacle avoidance instructions.

[0064] The drive execution unit is used to perform a safety assessment on obstacle avoidance commands. If the assessment is successful, the obstacle avoidance commands are converted into submersible control signals to drive the submersible to perform the avoidance.

[0065] A third aspect of this invention provides an electronic device, which includes:

[0066] At least one processor; and a memory communicatively connected to the at least one processor; wherein,

[0067] The memory stores instructions that can be executed by at least one processor, such that the at least one processor can perform the method proposed in the first aspect of the present invention.

[0068] A fourth aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method as described in the first aspect of the present invention. Attached Figure Description

[0069] Figure 1 This is a schematic diagram of the electronic device structure of the hardware operating environment involved in the embodiments of the present invention;

[0070] Figure 2 This is a flowchart illustrating the steps of an image recognition-based obstacle recognition method for submersibles provided in an embodiment of the present invention.

[0071] Figure 3This is a schematic diagram of the functional units of a submersible obstacle recognition device based on image recognition provided in an embodiment of the present invention. Detailed Implementation

[0072] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0073] The present invention will be further described below with reference to the accompanying drawings.

[0074] Reference Figure 1 , Figure 1 This is a schematic diagram of the electronic device structure of the hardware operating environment involved in the embodiments of the present invention.

[0075] like Figure 1 As shown, the electronic device may include: a processor 1001, such as a central processing unit (CPU), a communication bus 1002, a user interface 1003, a network interface 1004, and a memory 1005. The communication bus 1002 is used to enable communication between these components. The user interface 1003 may include a display screen or an input unit such as a keyboard; optionally, the user interface 1003 may also include a standard wired interface or a wireless interface. The network interface 1004 may optionally include a standard wired interface or a wireless interface (such as a Wi-Fi interface). The memory 1005 may be a high-speed random access memory (RAM) or a stable non-volatile memory (NVM), such as a disk drive. The memory 1005 may also optionally be a storage device independent of the aforementioned processor 1001.

[0076] Those skilled in the art will understand that Figure 1 The structure shown does not constitute a limitation on the electronic device and may include more or fewer components than shown, or combine certain components, or have different component arrangements.

[0077] like Figure 1 As shown, the memory 1005, which serves as a storage medium, may include an operating device, a data storage module, a network communication module, a user interface module, and electronic programs.

[0078] exist Figure 1 In the electronic device shown, the network interface 1004 is mainly used for data communication with the network server; the user interface 1003 is mainly used for data interaction with the user; the processor 1001 and the memory 1005 in the electronic device of the present invention can be set in the electronic device. The electronic device calls the image recognition-based submersible obstacle recognition device stored in the memory 1005 through the processor 1001 and executes the image recognition-based submersible obstacle recognition method provided in the embodiment of the present invention.

[0079] Reference Figure 2 The present invention provides an image recognition-based obstacle recognition method for submersibles, the method comprising:

[0080] S201: On the central server, build obstacle recognition and obstacle avoidance models, and deploy the obstacle recognition and obstacle avoidance models to all submersibles;

[0081] S202: Preprocess the multi-dimensional sensor data collected by the submersible, and input the preprocessed multi-dimensional sensor data into the obstacle recognition model to perform obstacle recognition and obtain the obstacle recognition result;

[0082] S203: Input the obstacle recognition results into the obstacle avoidance model to generate obstacle avoidance decisions, obtain the obstacle avoidance decisions, and generate corresponding obstacle avoidance instructions;

[0083] S204: Conduct a safety assessment of the obstacle avoidance command. If the assessment is successful, convert the obstacle avoidance command into a submersible control signal to drive the submersible to avoid the obstacle.

[0084] The technical solution provided in this application has at least the following beneficial effects:

[0085] By monitoring multi-dimensional sensor data, combining the detail advantages of optical images with the long-range, all-weather advantages of sonar images, the limitations of single sensors are overcome, significantly improving obstacle recognition accuracy and recall under various water quality and lighting conditions. A reinforcement learning-based obstacle avoidance model replaces traditional manual rules, enabling the submersible to autonomously learn complex avoidance strategies through interaction with the environment, enhancing its intelligence. Its decision-making is no longer a rigid "if-then" logic, but rather generates smooth, efficient, and human-driver-like intelligent avoidance behavior based on the distribution of multiple obstacles, its own state, and mission objectives. An obstacle avoidance system constructed using image recognition algorithms is employed. Unlike traditional image processing methods, deep learning models can automatically learn and extract deep, abstract features of obstacles from massive amounts of data, rather than relying on simple, manually designed features (such as edges and colors). This allows the model to effectively distinguish complex underwater backgrounds (such as seaweed, rocks, and shadows) from real obstacles. It is also highly robust to interference from changes in lighting, water scattering, and partial occlusion, thereby raising the recognition accuracy to a new level. Furthermore, it introduces an independent safety assessment process to perform secondary verification of decisions, effectively preventing collisions caused by model misjudgments or extreme situations, and providing double protection for the safe operation of submersibles.

[0086] In one alternative implementation, an obstacle recognition model and an obstacle avoidance model are built on a central server and deployed to all submersibles, including:

[0087] S2011: On the central server, an initial obstacle recognition model is built using an image recognition algorithm, an initial obstacle avoidance model is built using an intelligent decision-making algorithm, and the initial obstacle recognition model and the initial obstacle avoidance model are deployed to submersibles in different scenarios.

[0088] S2012: In each scene of the submersible, based on the collected first scene dataset, the initial obstacle recognition model is trained for the scene, and the first optimization algorithm is introduced to optimize the hyperparameters of the obstacle recognition model during the scene training process, so as to obtain the scene obstacle recognition model and several scene obstacle recognition results.

[0089] S2013: Based on the obstacle recognition results of several scenarios, the corresponding submersible status of the scenarios, and the target information of the scenarios, a second scenario dataset is constructed. Based on the second scenario dataset, the initial obstacle avoidance model is trained in the scenarios. A second optimization algorithm is introduced to optimize the action selection mechanism of the obstacle avoidance model, and the obstacle avoidance model of the scenario is obtained.

[0090] S2014: Extract the model update amount and scene data amount of the obstacle recognition model and the obstacle avoidance model of each submersible scene, and update the initial obstacle recognition model and the initial obstacle avoidance model of the central server according to all model update amounts and scene data amounts to obtain the global obstacle recognition model and the global obstacle avoidance model.

[0091] S2015: Deploy the global obstacle recognition model and the global obstacle avoidance model to all submersibles;

[0092] In this embodiment, the central server redeploys the updated global model to all submersibles, replacing the old local model. This completes one learning cycle. In new missions, the submersibles will begin a new round of local optimization based on the more powerful global model, forming a closed loop of continuous learning and evolution.

[0093] It is worth noting that by adopting a federated learning architecture, each submersible does not need to upload raw sensitive data, but only needs to upload model updates. Under the coordination of the central server, they can learn together and share wisdom. This not only protects data privacy and reduces communication costs, but more importantly, it achieves a synergistic evolution effect of "single-point breakthrough, group benefit". The intelligence level of the entire submersible group can be continuously and rapidly improved with the accumulation of mission experience.

[0094] In one optional implementation, the model update amount and scene data amount of the obstacle recognition model and obstacle avoidance model of each submersible scene are extracted. Based on all model update amounts and scene data amounts, the initial obstacle recognition model and initial obstacle avoidance model of the central server are updated to obtain the global obstacle recognition model and global obstacle avoidance model, including:

[0095] S20141: Extract the model update amount and scene data amount of the obstacle recognition model and obstacle avoidance model of each submersible scene, and send them to the central server.

[0096] S20142: On the central server, the importance weight of each submersible is generated based on the proportion of the scene data volume of each submersible to the total scene data volume of all submersibles.

[0097] The formula is:

[0098]

[0099] In the formula, Submersible generated based on scene data volume k The importance of weight; For submersible k,j The amount of scene data; k,jFor submersible indication; K The number of submersibles;

[0100] S20143: Based on the importance weight of each submersible and the corresponding model update amount, update the model parameters of the initial obstacle recognition model and the initial obstacle avoidance model to obtain the global model parameters.

[0101] The formula is:

[0102]

[0103] In the formula, Number of iterations t+ 1, t Global model parameters; For the first The amount of model parameter updates for each submersible; t This represents the current iteration number; To the number of iterations t The The important weight of each submersible; For submersible indication;

[0104] S20144: Based on the global model parameters, update the initial obstacle recognition model and the initial obstacle avoidance model of the central server to obtain the global obstacle recognition model and the global obstacle avoidance model.

[0105] In one alternative implementation, the obstacle recognition model is constructed based on the EfficientNet-Adaptive Feature Pyramid Network (AFPN) algorithm, and the obstacle recognition model includes a shared backbone network constructed based on the EfficientNet algorithm, a neck network constructed based on the AFPN algorithm, and a multi-task unified detection head. The multi-task unified detection head includes parallel type detection branches, position detection branches, size detection branches, and relative velocity detection branches. The first optimization algorithm is the Plant Root Growth Optimization (PRGO)-Particle Swarm Optimization (PSO) algorithm.

[0106] In this embodiment, the shared backbone network is used to efficiently extract multi-scale visual features from the preprocessed multi-dimensional sensor data; the neck network adopts an adaptive feature pyramid network to perform weighted fusion of feature maps at different levels extracted by the shared backbone network, thereby enhancing the ability to represent obstacles of different sizes; the multi-task unified detection head contains four parallel branches: a type detection branch (for classification, such as "reef", "fish", "net cage", etc.), a position detection branch (for outputting bounding boxes), a size detection branch (for estimating the actual size of the obstacle), and a relative velocity detection branch (for calculating the velocity vector of the obstacle relative to the submersible).

[0107] The obstacle avoidance model is constructed based on the Proximal Policy Optimization (PPO) algorithm, and the obstacle avoidance model includes an agent and an experience replay pool. The second optimization algorithm is the Ant Colony Optimization (ACO) algorithm.

[0108] The state space of the intelligent agent includes: obstacle recognition results (obstacle type, position, size, relative velocity), the submersible's own state (three-dimensional spatial position, velocity components of the submersible in three-dimensional space, three-axis attitude angles of the submersible (roll, pitch, yaw)), and target point information (straight-line distance from the submersible to the target point, horizontal azimuth angle of the target point in the submersible's body coordinate system, and pitch angle of the target point in the submersible's body coordinate system).

[0109] To simplify processing, we encode the information of the N most important obstacles (sorted by distance or threat level). Each obstacle includes the obstacle type, the relative distance between the obstacle and the submersible, the horizontal azimuth angle of the obstacle in the submersible's body coordinate system, the pitch angle of the obstacle in the submersible's body coordinate system, and the relative size of the obstacle (e.g., the logarithm of the pixel area it occupies in the image).

[0110] The action space is a three-dimensional continuous action: forward thrust, lateral thrust, and vertical thrust; the range of the forward main thrust is [-1, 1], where -1 represents the maximum reverse thrust and 1 represents the maximum forward thrust; the range of the lateral thrust is [-1, 1], used for translation and turning; and the range of the vertical thrust is [-1, 1], used for diving and surfacing.

[0111] The formula for the reward function:

[0112]

[0113] In the formula, for Momentary rewards; For time indication; As a reward weight; These are respectively: target approach reward, safety reward for moving away from obstacles, and smooth movement reward;

[0114] Formula for rewarding the approach of a target:

[0115]

[0116] In the formula, for The distance to the target at any given time is used as the reward; if the distance decreases, the reward is positive; otherwise, the reward is negative. The normalization coefficient is used to prevent the reward value from being too large;

[0117] Formula for safety reward:

[0118]

[0119] In the formula, The minimum distance to all obstacles at the current moment; As a preset safe distance threshold, when > When the distance increases, the reward is positive and increases slowly; when... < At that time, the reward decreased sharply, and in Less than the collision distance At that time, a huge negative reward (such as -100) is given.

[0120] Formula for motion smoothing reward:

[0121]

[0122] In the formula, for The constant movement, and the dramatic changes in this penalty action, encourage the submersible to navigate smoothly, conserve energy, and avoid mechanical wear and tear.

[0123] In one optional implementation, for each scene in the submersible, based on the collected first scene dataset, an initial obstacle recognition model is trained for the scene, and a first optimization algorithm is introduced to optimize the hyperparameters of the obstacle recognition model during scene training, resulting in a scene obstacle recognition model and several scene obstacle recognition results, including:

[0124] S20121: In each scene of the submersible, based on the first scene dataset with the first prediction label (i.e. the true type, position, size, and relative speed of the obstacle) collected, the initial obstacle recognition model is trained in the scene, and the PRGO-PSO algorithm is introduced in the scene training process.

[0125] S20122: Encode the hyperparameters of the obstacle recognition model into individual vectors of the PRGO algorithm, and set the fitness function, PRGO population parameters, and maximum number of iterations;

[0126] In this embodiment, hyperparameters include the learning rate, batch size, and weight decay of the obstacle recognition model;

[0127] The fitness function is the function that calculates the accuracy of the model on the validation set;

[0128] S20123: Based on the PRGO population parameters, the initial PRGO population is obtained by initializing using the Tent chaotic mapping sequence; each PRGO individual in the PRGO population corresponds to a candidate hyperparameter.

[0129] The formula is:

[0130]

[0131] In the formula, For the initial PRGO population, the first i One initial PRGO individual; For the first i One chaotic variable; These are the upper and lower bounds of the search space; i For PRGO individual indicators;

[0132]

[0133] In the formula, For the first i- One chaotic variable; compared with random initialization, chaotic initialization can ensure that the population is evenly distributed in the solution space, thus enhancing diversity.

[0134] S20124: Based on the first scene dataset, use the fitness function to obtain the fitness value of each initial PRGO individual, and take the initial PRGO individual with the best fitness value as the optimal solution;

[0135] S20125: Based on the fiber root growth behavior and lateral root growth behavior of the PRGO algorithm, the initial PRGO population is iteratively updated to obtain the updated first PRGO population.

[0136] The formula is:

[0137]

[0138] In the formula, Number of iterations t+ The first fiber root growth obtained i The first updated PRGO individual; The convergence factor; Number of iterations t The optimal solution; PRGO individuals selected randomly; t This represents the current iteration number;

[0139]

[0140] In the formula, These are the maximum and minimum values ​​of the convergence factor; This represents the maximum number of iterations. t This represents the current iteration number; , To adjust the parameters; It is the hyperbolic tangent function;

[0141]

[0142] In the formula, Number of iterations t+ The lateral root growth of 1 obtained the first i The first updated PRGO individual; for Levy Distribute random numbers; b for Levy Step length, and b ∈[1,2]; Number of iterations t The mean of the fiber roots;

[0143]

[0144] In the formula, Number of iterations t The first fiber root growth obtained The first updated PRGO individual; For PRGO individual indicators; The number of PRGO individuals;

[0145] In this embodiment, the trend of plant roots growing towards water and nutrients (i.e., better fitness) is simulated to conduct global exploration and local development of the population.

[0146] S20126: Based on the improved primary root growth behavior of the PSO mechanism of the PRGO algorithm, the updated first PRGO population is iteratively updated to obtain the updated second PRGO population.

[0147] The particle swarm optimization algorithm is introduced to improve the principal root growth behavior, and the formula is as follows:

[0148]

[0149] In the formula, Number of iterations t+ 1 of i The particle velocity of the first updated PRGO individual; This is the inertia weight, with a value range of 0.4-0.9; Use random numbers in the range [0,1] to increase the randomness of the search and avoid getting trapped in local optima; This is the learning factor, with a value range of 1.5-2.0; For each individual optimal solution, record the particle's own optimal solution to guide the local search; This is the globally optimal solution, corresponding to the optimal solution of the first PRGO population in the last update, guiding the population to converge toward the optimal region; Number of iterations t The i The first updated PRGO individual;

[0150]

[0151] In the formula, Number of iterations t+ The first root growth behavior obtained i A new second PRGO individual;

[0152] In this embodiment, the concepts of particle velocity, individual optimality, and global optimality from the PSO algorithm are introduced to guide the principal root to perform a more targeted search towards the optimal solution, thereby enhancing the convergence speed and accuracy.

[0153] S20127: Based on the first scene dataset, use the fitness function to obtain the fitness value of each updated second PRGO individual, and update the updated second PRGO individual with the best fitness value as the optimal solution;

[0154] S20128: When the number of iterations reaches the maximum number of iterations or the fitness value of the optimal solution meets the requirements, terminate the iterative update of the PRGO population and output the optimal solution of the current iteration;

[0155] S20129: Decode the individual vectors of the PRGO individuals corresponding to the optimal solution to obtain the optimized hyperparameters, and obtain the optimized obstacle recognition model based on the optimized hyperparameters;

[0156] In this embodiment, the PRGO-PSO hybrid optimization algorithm is used to automatically find the hyperparameters. The PRGO-PSO algorithm combines the global exploration capability of the plant root algorithm and the local development capability of the particle swarm algorithm, and enhances the population diversity through Tent chaotic mapping initialization. It can quickly and stably find the optimal combination of hyperparameters, which significantly improves the model's recognition performance in specific scenarios.

[0157] S201210: Input the first scene dataset into the optimized obstacle recognition model for fine-tuning to obtain the scene obstacle recognition model and several scene obstacle recognition results.

[0158] In one optional implementation, a second scene dataset is constructed based on obstacle recognition results from several scenes, the corresponding submersible state in the scene, and scene target information. The initial obstacle avoidance model is then trained on this second scene dataset. A second optimization algorithm is introduced to optimize the action selection mechanism of the obstacle avoidance model, resulting in a scene-specific obstacle avoidance model, including:

[0159] S20131: Associate the obstacle recognition results of each scene, the corresponding scene submersible status, and the scene target information, and add a second prediction label (i.e., the optimal avoidance action in this state) to obtain the second scene dataset.

[0160] S20132: Based on the second scenario dataset, the initial obstacle avoidance model is trained in the scenario, and the pheromone mechanism of the ACO algorithm is introduced during the scenario training process.

[0161] In this embodiment, the idea of ​​swarm intelligence optimization is introduced into reinforcement learning. By guiding strategy exploration through a virtual pheromone map, the traditional "blind trial and error" mode of reinforcement learning is completely changed. The agent can make decisions on the shoulders of "historical experience", and the learning speed and strategy quality are improved by orders of magnitude.

[0162] S20133: Set up a virtual pheromone map and pheromone guide for the initial obstacle avoidance model, and set dynamic update rules for the virtual pheromone map based on the reward mechanism of the obstacle avoidance model to obtain an optimized obstacle avoidance model.

[0163] In this embodiment, a three-dimensional voxel grid is used to represent the virtual pheromone map, the resolution and range of which are consistent with the mission planning area of ​​the submersible. Figure 1 Each voxel stores a floating-point number representing the pheromone concentration at that point. The higher the concentration, the more "worth exploring" or "a proven safe and high-quality path" the area is.

[0164] S20134: Input the second scene dataset into the optimized obstacle avoidance model for fine-tuning to obtain the obstacle avoidance model for the scene.

[0165] In one alternative implementation, the multidimensional sensor data includes optical image data, sonar scan data, and the submersible's own pose data;

[0166] Obstacle identification results include the type, location, size, and relative speed of the obstacle.

[0167] In one optional implementation, the multidimensional sensor data collected by the submersible is preprocessed, and the preprocessed multidimensional sensor data is input into an obstacle recognition model for obstacle recognition, resulting in obstacle recognition results, including:

[0168] S2021: Using the multi-source sensors of the submersible, multi-dimensional sensor data is collected, and the multi-dimensional sensor data is spatiotemporally synchronized, denoised, enhanced, and spatially aligned to obtain pre-processed multi-dimensional sensor data, which is then input into the obstacle recognition model.

[0169] In this embodiment, time-space synchronization is achieved using timestamps and spatial transformation matrices; noise reduction is performed using filtering algorithms (such as median filtering and Gaussian filtering); image enhancement is achieved using methods such as histogram equalization and contrast stretching; and spatial alignment is achieved by converting data from different sensors (such as cameras and sonar) to the same coordinate system.

[0170] S2022: Using a shared backbone network of an obstacle recognition model, extract multi-scale features from preprocessed multi-dimensional sensor data;

[0171] S2023: Using the neck network of the obstacle recognition model, multi-scale features are weighted and fused to obtain fused features;

[0172] S2024: Based on the fusion features, use a multi-task unified detection head to generate obstacle recognition results.

[0173] In one optional implementation, the obstacle recognition result is input into the obstacle avoidance model to generate an obstacle avoidance decision, thereby obtaining the obstacle avoidance decision and generating corresponding obstacle avoidance instructions, including:

[0174] S2031: Collect the submersible status and target information corresponding to the obstacle recognition results, and input the obstacle recognition results, submersible status and target information into the obstacle avoidance model;

[0175] S2032: Initialize the virtual pheromone map of the obstacle avoidance model based on the submersible's mission planning map;

[0176] In this embodiment, several areas of the virtual pheromone map are initialized:

[0177] Known safe zones: If there is prior knowledge (such as planned waterways or open waters confirmed in historical missions), the voxel values ​​corresponding to these zones are initialized to a higher base value (e.g., 0.7).

[0178] Unknown regions: The voxel values ​​of all other unknown regions are initialized to a lower default value (e.g., 0.3);

[0179] Known danger zones: If there are known restricted or high-risk zones (such as military restricted zones or areas with dense fishing nets), initialize their voxel values ​​to 0;

[0180] S2033: Write the obstacle recognition results, the submersible status and the scene target information into the state space of the agent, randomly extract a number of historical experiences from the experience replay pool, and write the action set of the historical experiences into the action space of the agent.

[0181] Based on the state space, the agent's policy network is used to generate an initial probability distribution in the action space, and the initial probability distribution is discretized to generate the corresponding candidate action set.

[0182] In this embodiment, the "original optimal action" output by the Actor network is considered. (i.e., the mean of a Gaussian distribution), perform gridding or random sampling to generate the corresponding candidate action set. ,in, There are M candidate actions;

[0183] For example, if Given [0.8, 0.1, -0.2], the following candidates can be generated:

[0184] = [0.8, 0.1, -0.2] (Original action)

[0185] = [0.8, 0.3, -0.2] (slightly to the right)

[0186] = [0.8, -0.1, -0.2] (slightly to the left)

[0187] And so on, generating M (e.g., 9) candidate actions that cover the area surrounding the original action;

[0188] S2034: Using a pheromone guide, predict the future position of each candidate action in the candidate action set, and query the pheromone concentration of the future position on the virtual pheromone map;

[0189] In this embodiment, based on a simplified kinematic model of the submersible, a pheromone guide is used to predict the outcome of a predicted action. A short time Δ t The possible future location of the submersible after (e.g., 2 seconds). ,in, For action indication quantity, For three-dimensional coordinates; the formula for the kinematic model is:

[0190]

[0191] In the formula, For future location; Current position; Current speed; For predicting time; To predict actions;

[0192] In virtual pheromone map In the middle, search The pheromone concentration value of the voxel ,if If the pheromone concentration is outside the map area, an extremely low pheromone concentration value is assigned.

[0193] S2035: Based on the pheromone concentration, calculate the pheromone bias weight of each candidate action in each candidate action set, and generate the final selection probability of the candidate action set based on the pheromone bias weight.

[0194] In this embodiment, for each candidate action Calculate a pheromone-guided pheromone bias weight This weight reflects the "attractiveness" of the future location to which the action is directed;

[0195] The formula is:

[0196]

[0197] In the formula, For guiding strength coefficient, >0, The larger the size, the stronger the guiding effect of the pheromone; Candidate actions pheromone concentration values;

[0198] The Actor network originally consisted of each An original probability was assigned. (Calculated from the Gaussian probability density function), now, using weights By biasing this probability, we obtain the final selection probability. ;

[0199] The formula is:

[0200]

[0201] In the formula, For the first Each pheromone bias weight and original probability; For the action indication; this formula ensures that the probability distribution after biasing still sums to 1;

[0202] S2036: Final selection probability The probability distribution is transformed into a final probability distribution, and based on the final probability distribution, the action to be executed, i.e. the obstacle avoidance decision, is selected, and the corresponding obstacle avoidance command is generated.

[0203] S2037: Based on the reward function, the agent generates a reward value for the action performed, and updates the virtual pheromone map according to the reward value based on the dynamic update rule, thus obtaining an updated virtual pheromone map;

[0204] In this embodiment, the dynamic update rules include a pheromone evaporation mechanism and a positive and negative feedback mechanism;

[0205] Pheromones volatilization mechanism:

[0206] To simulate the evaporation characteristics of pheromones in nature over time and prevent old information from excessively influencing new decisions, global evaporation is performed on all voxels of the virtual pheromone map, and the voxels after evaporation are used as the initial voxels for updating the virtual pheromone map.

[0207] The formula is:

[0208]

[0209] In the formula, The evaporation rate is a very small positive number (e.g., 0.01). Voxels representing virtual pheromone maps before and after evaporation;

[0210] Positive and negative feedback mechanisms:

[0211] Positive feedback (reward deposition):

[0212] When the agent receives a significant positive reward (For example, This indicates a successful evasion or a significant approach to the target, triggering positive feedback, in which... For time indication quantity, The positive reward threshold;

[0213] Deposition location: in state Corresponding submersible location Deposition occurs within a spherical area around the food source, mimicking the pheromones left by ants on their return path after finding food.

[0214] Sediment volume: Sediment volume With rewards They are directly proportional, and the formula is:

[0215]

[0216] In the formula, The deposition coefficient;

[0217] Sedimentation Algorithm: Traversal The surrounding radius is All voxels are deposited in a Gaussian distribution according to their distance:

[0218]

[0219] In the formula, Voxels for virtual pheromone maps before and after deposition; For distance; The standard deviation is 1; this makes the pheromone concentration highest at the center of the path and gradually decreases towards the edges.

[0220] Negative feedback (danger marker):

[0221] When the agent receives a huge negative reward (For example, This usually means a collision or extreme danger), triggering negative feedback, in which... The negative reward threshold;

[0222] Mark location: In states that result in negative rewards Corresponding submersible location Mark a spherical area around it;

[0223] Labeling amount: Directly set the voxel value of the region to a very low value, or multiply it by a decay factor (such as 0.1).

[0224] The formula is:

[0225]

[0226] In the formula, The evaporation rate is a very small positive number (e.g., 0.01). Voxels representing virtual pheromone maps before and after decay; this is equivalent to leaving "warning pheromones" in danger zones to warn future exploration activities to stay away from these areas.

[0227] The positive and negative feedback mechanism of pheromones enables submersibles to learn online and dynamically adapt to environmental changes, demonstrating a strong ability to adapt to unknown obstacles and dynamic threats.

[0228] S2038: Store the real-time experience generated by obstacle avoidance decision into the experience replay pool, and continuously train the obstacle avoidance model based on the real-time experience and some historical experience to obtain an updated obstacle avoidance model.

[0229] In this embodiment, through online learning and continuous training mechanisms, the submersible can continuously optimize its own model using data accumulated in actual tasks, achieving self-evolution and greatly improving the autonomy and adaptability of the submersible in long-term deployment missions.

[0230] In one optional implementation, a safety assessment is performed on the obstacle avoidance command. If the assessment passes, the obstacle avoidance command is converted into a submersible control signal to drive the submersible to perform the avoidance maneuver, including:

[0231] S2041: Perform a safety boundary check on the obstacle avoidance command. If the safety boundary check passes, proceed to the next step; otherwise, issue a safety alarm and block the obstacle avoidance command.

[0232] In this embodiment, it checks whether the command exceeds the physical limits of the submersible, such as the maximum turning angle, maximum pitch angle, maximum thruster power, etc. If it exceeds these limits, it is determined to be an unsafe command, a safety alarm is issued, and the execution of the command is prevented. It can revert to the safe default action (such as hovering or emergency ascent).

[0233] S2042: Perform dynamic trajectory prediction and collision risk assessment on obstacle avoidance commands. If the obtained dynamic trajectory has no collision risk, proceed to the next step; otherwise, issue a safety alarm and block the obstacle avoidance command.

[0234] In this embodiment, based on the current command and the dynamic model of the submersible, the motion trajectory within a certain period of time in the future is predicted. At the same time, combined with the obstacle recognition results and their relative speed, the future trajectory of the obstacle is predicted. The two trajectories are subjected to collision detection. If a collision risk is predicted, it is determined to be an unsafe command, and an alarm is issued and execution is prevented.

[0235] S2043: Based on the historical obstacle avoidance command from the previous moment, perform motion smoothing on the current obstacle avoidance command and convert it into the corresponding submersible control signal to drive the submersible to avoid obstacles.

[0236] In this embodiment, in order to prevent sudden changes in commands from impacting the mechanical structure, the system combines the historical commands from the previous moment and performs smooth filtering on the current command to generate a continuous and smooth control signal. Finally, the control signal is sent to the submersible's thrusters, servo motors and other actuators to drive the submersible to complete the evasive maneuver.

[0237] The smoothing filter uses a weighted average, and the formula is:

[0238]

[0239] In the formula, The weighted average thrust value is the thrust value corresponding to the command, the thrust value corresponding to the current command, and the thrust value corresponding to the historical command at the previous moment.

[0240] This invention also provides an image recognition-based obstacle recognition device for submersibles, referring to... Figure 3 The diagram shows a functional unit diagram of a submersible obstacle recognition device 300 based on image recognition according to the present invention. The device may include the following units:

[0241] The model deployment unit 301 is used to build obstacle recognition models and obstacle avoidance models on the central server, and to deploy the obstacle recognition models and obstacle avoidance models to all submersibles;

[0242] The obstacle recognition unit 302 is used to preprocess the multi-dimensional sensor data collected by the submersible, and input the preprocessed multi-dimensional sensor data into the obstacle recognition model to perform obstacle recognition and obtain the obstacle recognition result.

[0243] The obstacle avoidance unit 303 is used to input the obstacle recognition result into the obstacle avoidance model, generate obstacle avoidance decision, obtain the obstacle avoidance decision, and generate the corresponding obstacle avoidance instruction.

[0244] The drive execution unit 304 is used to perform a safety assessment on the obstacle avoidance command. If the assessment is successful, the obstacle avoidance command is converted into a submersible control signal to drive the submersible to avoid the obstacle.

[0245] Based on the same inventive concept, another embodiment of the present invention provides an electronic device, including a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus.

[0246] Memory, used to store computer programs;

[0247] When the processor executes the program stored in the memory, it implements the image recognition-based obstacle recognition method for submersibles of the present invention.

[0248] The communication bus mentioned above can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. This communication bus can be divided into address bus, data bus, control bus, etc. For ease of representation, only one thick line is used in the diagram, but this does not indicate that there is only one bus or one type of bus. The communication interface is used for communication between the aforementioned terminal and other devices. The memory can include Random Access Memory (RAM) or non-volatile memory, such as at least one disk storage device. Optionally, the memory can also be at least one storage device located remotely from the aforementioned processor.

[0249] The processors mentioned above can be general-purpose processors, including central processing units (CPUs), network processors (NPs), etc.; they can also be digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0250] Furthermore, to achieve the above objectives, embodiments of the present invention also propose a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the image recognition-based obstacle recognition method for submersibles according to embodiments of the present invention.

[0251] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, apparatus, or computer program products. Therefore, embodiments of the present invention can take the form of entirely hardware embodiments, entirely software embodiments, or embodiments combining software and hardware aspects. Furthermore, embodiments of the present invention can take the form of computer program products implemented on one or more computer-usable hardware devices (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0252] The embodiments of the present invention are described with reference to flowchart illustrations and / or block diagrams of methods, terminal devices (apparatus), and computer program products according to embodiments of the invention. It should be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing terminal device to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing terminal device, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0253] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing terminal device to operate in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0254] These computer program instructions can also be loaded onto a computer or other programmable data processing terminal equipment, causing a series of operational steps to be performed on the computer or other programmable terminal equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable terminal equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0255] Finally, it should be noted that in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. "And / or" indicates that either one or both can be chosen. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or terminal device that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or terminal device. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or terminal device that includes the element.

[0256] The above are merely specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and these modifications or substitutions should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for obstacle recognition in submersibles based on image recognition, characterized in that, The method includes: On the central server, obstacle recognition and obstacle avoidance models are built and deployed to all submersibles, including: On the central server, an initial obstacle recognition model is built using image recognition algorithms, an initial obstacle avoidance model is built using intelligent decision-making algorithms, and the initial obstacle recognition model and the initial obstacle avoidance model are deployed to submersibles in different scenarios. In each scene of the submersible, based on the collected first scene dataset, the initial obstacle recognition model is trained for the scene, and the first optimization algorithm is introduced to optimize the hyperparameters of the obstacle recognition model during the scene training process, so as to obtain the scene obstacle recognition model and several scene obstacle recognition results. The obstacle recognition model is constructed based on the EfficientNet-AFPN algorithm, and includes a shared backbone network constructed based on the EfficientNet algorithm, a neck network constructed based on the AFPN algorithm, and a multi-task unified detection head. The multi-task unified detection head includes parallel type detection branches, position detection branches, size detection branches, and relative velocity detection branches. The first optimization algorithm is the PRGO-PSO algorithm. include: In each scene of the submersible, the initial obstacle recognition model is trained on the scene based on the collected first scene dataset with the first prediction label, and the PRGO-PSO algorithm is introduced during the scene training process. The hyperparameters of the obstacle recognition model are encoded into individual vectors of the PRGO algorithm, and the fitness function, PRGO population parameters, and maximum number of iterations are set. Based on the PRGO population parameters, the initial PRGO population is obtained by initializing using the Tent chaotic mapping sequence; each PRGO individual in the PRGO population corresponds to a candidate hyperparameter. Based on the first scenario dataset, the fitness function is used to obtain the fitness value of each initial PRGO individual, and the initial PRGO individual with the best fitness value is taken as the optimal solution. Based on the fiber root growth behavior and lateral root growth behavior of the PRGO algorithm, the initial PRGO population is iteratively updated to obtain the updated first PRGO population. Based on the improved primary root growth behavior of the PSO mechanism of the PRGO algorithm, the updated first PRGO population is iteratively updated to obtain the updated second PRGO population. Based on the first scenario dataset, the fitness function is used to obtain the fitness value of each updated second PRGO individual, and the updated second PRGO individual with the best fitness value is updated as the optimal solution; When the number of iterations reaches the maximum number of iterations or the fitness value of the optimal solution meets the requirements, the iterative update of the PRGO population is terminated, and the optimal solution of the current iteration is output. The individual vectors of the PRGO individuals corresponding to the optimal solution are decoded to obtain the optimized hyperparameters, and the optimized obstacle recognition model is obtained based on the optimized hyperparameters. The first scene dataset is input into the optimized obstacle recognition model for fine-tuning, resulting in the scene obstacle recognition model and several scene obstacle recognition results; Based on the obstacle recognition results of several scenarios, the corresponding submersible status of the scenarios, and the target information of the scenarios, a second scenario dataset is constructed. Based on the second scenario dataset, the initial obstacle avoidance model is trained on the scenarios. A second optimization algorithm is introduced to optimize the action selection mechanism of the obstacle avoidance model, thereby obtaining the obstacle avoidance model of the scenarios. The obstacle avoidance model is constructed based on the PPO algorithm, and the obstacle avoidance model includes an agent and an experience replay pool. The second optimization algorithm is the ACO algorithm. include: The obstacle recognition results for each scene, the corresponding submersible status, and the scene target information are associated, and a second prediction label is added to obtain the second scene dataset. Based on the second scenario dataset, the initial obstacle avoidance model was trained in the scenario, and the pheromone mechanism of the ACO algorithm was introduced during the scenario training process. A virtual pheromone map and pheromone guide are set up for the initial obstacle avoidance model, and a dynamic update rule is set for the virtual pheromone map based on the reward mechanism of the obstacle avoidance model to obtain an optimized obstacle avoidance model. The second scene dataset is input into the optimized obstacle avoidance model for fine-tuning to obtain the obstacle avoidance model for the scene. Extract the model update amount and scene data amount of the obstacle recognition model and the obstacle avoidance model of each submersible scene. Based on the update amount of all models and the scene data amount, update the initial obstacle recognition model and the initial obstacle avoidance model of the central server to obtain the global obstacle recognition model and the global obstacle avoidance model. Deploy the global obstacle recognition model and the global obstacle avoidance model to all submersibles; The multidimensional sensor data collected by the submersible is preprocessed, and the preprocessed multidimensional sensor data is input into the obstacle recognition model to perform obstacle recognition and obtain the obstacle recognition result. The obstacle recognition results are input into the obstacle avoidance model to generate obstacle avoidance decisions, and corresponding obstacle avoidance instructions are generated. A safety assessment is performed on the obstacle avoidance command. If the assessment is successful, the obstacle avoidance command is converted into a submersible control signal to drive the submersible to avoid the obstacle.

2. The underwater obstacle recognition method based on image recognition according to claim 1, characterized in that, The multidimensional sensor data includes optical image data, sonar scan data, and the submersible's own pose data; The obstacle identification results include the type, location, size, and relative speed of the obstacle.

3. The underwater obstacle recognition method based on image recognition according to claim 2, characterized in that, The multidimensional sensor data collected by the submersible is preprocessed, and the preprocessed multidimensional sensor data is input into the obstacle recognition model for obstacle recognition. The obstacle recognition results are then obtained, including: Using the multi-source sensors of the submersible, multi-dimensional sensor data is collected, and the multi-dimensional sensor data is spatiotemporally synchronized, denoised, enhanced, and spatially aligned to obtain pre-processed multi-dimensional sensor data, which is then input into the obstacle recognition model. Using a shared backbone network of an obstacle recognition model, multi-scale features are extracted from preprocessed multi-dimensional sensor data. The neck network of the obstacle recognition model is used to perform weighted fusion of multi-scale features to obtain fused features; Based on the fusion features, a multi-task unified detection head is used to generate obstacle recognition results.

4. The underwater obstacle recognition method based on image recognition according to claim 3, characterized in that, The obstacle recognition results are input into the obstacle avoidance model to generate obstacle avoidance decisions, resulting in corresponding obstacle avoidance instructions, including: Collect the submersible status and target information corresponding to the obstacle recognition results, and input the obstacle recognition results, submersible status and target information into the obstacle avoidance model; Based on the submersible's mission planning map, initialize the virtual pheromone map of the obstacle avoidance model; The obstacle recognition results, the submersible status, and the scene target information are written into the state space of the agent. Several historical experiences are randomly extracted from the experience replay pool, and the action set of the historical experiences is written into the action space of the agent. Based on the state space, the agent's policy network is used to generate an initial probability distribution in the action space, and the initial probability distribution is discretized to generate the corresponding candidate action set. Using a pheromone guide, predict the future location of each candidate action in the candidate action set, and query the pheromone concentration at the future location on a virtual pheromone map; Based on the pheromone concentration, calculate the pheromone bias weight for each candidate action in each candidate action set, and generate the final selection probability of the candidate action set based on the pheromone bias weight. The final selection probability is transformed into a final probability distribution, and based on the final probability distribution, the action to be executed, i.e. the obstacle avoidance decision, is selected and the corresponding obstacle avoidance instruction is generated. Based on the reward function, the agent generates reward values ​​for the actions performed, and updates the virtual pheromone map according to the reward values ​​based on the dynamic update rules, thus obtaining an updated virtual pheromone map. The real-time experience generated by obstacle avoidance decision-making is stored in the experience replay pool, and the obstacle avoidance model is continuously trained based on the real-time experience and some historical experience to obtain an updated obstacle avoidance model.

5. The underwater obstacle recognition method based on image recognition according to claim 4, characterized in that, A safety assessment is conducted on the obstacle avoidance command. If the assessment is successful, the obstacle avoidance command is converted into a submersible control signal to drive the submersible to perform the avoidance maneuver, including: The obstacle avoidance command is checked for safety boundaries. If the safety boundary check passes, the process proceeds to the next step; otherwise, a safety alarm is issued and the obstacle avoidance command is blocked. The obstacle avoidance command is dynamically trajectory predicted and collision risk assessed. If the obtained dynamic trajectory has no collision risk, the next step is initiated; otherwise, a safety alarm is issued and the obstacle avoidance command is blocked. Based on the historical obstacle avoidance commands from the previous moment, the current obstacle avoidance commands are processed for motion smoothing and converted into corresponding submersible control signals to drive the submersible to perform the avoidance maneuver.

6. A submersible obstacle recognition device based on image recognition, used to implement the submersible obstacle recognition method as described in any one of claims 1-5, characterized in that, The device includes: The model deployment unit is used to build obstacle recognition models and obstacle avoidance models on the central server, and to deploy the obstacle recognition models and obstacle avoidance models to all submersibles; The obstacle recognition unit is used to preprocess the multi-dimensional sensor data collected by the submersible, and input the preprocessed multi-dimensional sensor data into the obstacle recognition model to perform obstacle recognition and obtain the obstacle recognition result. The obstacle avoidance unit is used to input the obstacle recognition results into the obstacle avoidance model, generate obstacle avoidance decisions, obtain obstacle avoidance decisions, and generate corresponding obstacle avoidance instructions. The drive execution unit is used to perform a safety assessment on obstacle avoidance commands. If the assessment is successful, the obstacle avoidance commands are converted into submersible control signals to drive the submersible to perform the avoidance.

Images